diff --git a/.gitignore b/.gitignore
index 8d4ceaa811c04..54129af906fbd 100644
--- a/.gitignore
+++ b/.gitignore
@@ -124,6 +124,13 @@ torch/utils/benchmark/utils/valgrind_wrapper/callgrind.h
 torch/utils/benchmark/utils/valgrind_wrapper/valgrind.h
 torch/version.py
 minifier_launcher.py
+aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_fwd_d*
+aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd_d*
+aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd_convert*
+aten/src/ATen/native/transformers/hip/flash_attn/ck/fwd_blob*
+aten/src/ATen/native/transformers/hip/flash_attn/ck/bwd_blob*
+aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_fwd_api*
+aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd_api*
 # Root level file used in CI to specify certain env configs.
 # E.g., see .circleci/config.yaml
 env
diff --git a/aten/src/ATen/CMakeLists.txt b/aten/src/ATen/CMakeLists.txt
index ab95de5036bdc..874b45688d50b 100644
--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
@@ -183,6 +183,8 @@ if(USE_FLASH_ATTENTION)
           endif()
         endif()
         message(STATUS "USE_CK_FLASH_ATTENTION is set; building PyTorch with CK Flash Attention enabled")
+        message(STATUS "Generating CK kernel instances...")
+        add_subdirectory(native/transformers/hip/flash_attn/ck)
         file(GLOB flash_attention_hip_ck_hip "native/transformers/hip/flash_attn/ck/*.hip")
         list(APPEND native_transformers_hip_hip ${flash_attention_hip_ck_hip})
       endif()
diff --git a/aten/src/ATen/native/transformers/attention.cpp b/aten/src/ATen/native/transformers/attention.cpp
index b96787cbda5f2..7beb24b8cb6c6 100644
--- a/aten/src/ATen/native/transformers/attention.cpp
+++ b/aten/src/ATen/native/transformers/attention.cpp
@@ -74,7 +74,7 @@
 #include <ATen/ops/_safe_softmax_native.h>
 #include <ATen/ops/all.h>
 #endif
-
+#include <iostream>
 #include <ATen/native/nested/NestedTensorTransformerFunctions.h>
 namespace at::native {
 
@@ -741,6 +741,7 @@ Tensor scaled_dot_product_attention(
       if (attn_mask.has_value()) {
         attn_mask.value() = preprocess_mask(attn_mask.value(), query_, key, value);;
       }
+      //std::cout << "OUTERMOST Q SHAPE: " << query_.sizes() << std::endl;
       auto out_and_lse = at::_scaled_dot_product_efficient_attention(
           query_, key, value, attn_mask, compute_logsumexp, dropout_p, is_causal, scale);
       return std::get<0>(out_and_lse);
diff --git a/aten/src/ATen/native/transformers/cuda/attention.cu b/aten/src/ATen/native/transformers/cuda/attention.cu
index 7fe7ee7a1ba19..6d76653dcb3cb 100644
--- a/aten/src/ATen/native/transformers/cuda/attention.cu
+++ b/aten/src/ATen/native/transformers/cuda/attention.cu
@@ -82,6 +82,7 @@
 #include <ATen/native/transformers/hip/aotriton_adapter.h>
 #include <aotriton/flash.h>
 #include <aotriton/runtime.h>
+#include <ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h>
 #endif
 #endif
 
@@ -848,13 +849,25 @@ std::tuple<Tensor, Tensor, Tensor, Tensor> _scaled_dot_product_efficient_attenti
     std::optional<double> scale) {
   // Used for tracking usage statistics
   C10_LOG_API_USAGE_ONCE("torch.sdpa.mem_efficient_attention");
+  //std::cout << std::endl;
+  //std::cout << "what we want vvvvvvvvvvvvvv" << std::endl;
+  //std::cout << "MAX_SEQLEN_Q: " << b4_max_seqlen_batch_q << std::endl;
+  //std::cout << "MAX_SEQLEN_K: " << b4_max_seqlen_batch_k << std::endl;
+  //std::cout << "MAX_SEQLEN_V: " << b4_max_seqlen_batch_v << std::endl;
+  //std::cout << "^^^^^^^^^^^^^^^^^^^^^^^^^^^^" << std::endl;
   // Query -> Query(Batch x Q_seq_len x Num_heads x Dim_per_head)
   // Key   -> Key(Batch x KV_seq_len x Num_heads x Dim_per_head)
   // Value -> Value(Batch x KV_seq_len x  Num_heads x Dim_per_head)
+  std::cout << std::endl;
+  std::cout << "sdpa_ef" << std::endl;
+  std::cout << "q.sizes  : " << query.sizes() << std::endl;
   Tensor q_t = query.transpose(1, 2);
   Tensor k_t = key.transpose(1, 2);
   Tensor v_t = value.transpose(1, 2);
 
+  std::cout << "q_t.sizes: " << q_t.sizes() << std::endl;
+
+  std::cout << "qagain.sizes: " << query.sizes() << std::endl;
   sdp::CustomMaskType custom_mask_type = is_causal
       ? sdp::CustomMaskType::CausalFromTopLeft
       : sdp::CustomMaskType::NoCustomMask;
@@ -1026,6 +1039,16 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, c10::SymInt, c10::SymInt> _efficient_
 // TODO In theory it is possible to compile with _CUDA_ARCH < 5.0 and run on a
 // machine that is >= 5.0. In practice, this is not a problem but since
 // this would avoid runtime architecture checks, we should look into it
+  const int64_t new_max_seqlen_batch_q = query.size(1);
+  const int64_t new_max_seqlen_batch_k = key.size(1);
+  const int64_t new_max_seqlen_batch_v = value.size(1);
+  std::cout << std::endl;
+  std::cout << "MEMORY_EFFICIENT VVVVVVVVVV" << std::endl;
+  std::cout << "MAX_SEQLEN_Q: " << new_max_seqlen_batch_q << std::endl;
+  std::cout << "MAX_SEQLEN_K: " << new_max_seqlen_batch_k << std::endl;
+  std::cout << "MAX_SEQLEN_V: " << new_max_seqlen_batch_v << std::endl;
+  std::cout << "^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^" << std::endl;
+  TORCH_CHECK(query.dim() == 4);
 
   TORCH_CHECK(query.dim() == 4);
   TORCH_CHECK(key.dim() == 4);
@@ -1128,97 +1151,145 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, c10::SymInt, c10::SymInt> _efficient_
 
 #ifdef USE_ROCM
   // ROCM Implementation
-  auto ret = aotriton::v2::flash::check_gpu(stream);
-  if (hipSuccess != ret) {
-      TORCH_CHECK(false,
-                  "[AOTriton] Accelerated SDPA only supports MI200/MI300X/Navi31 GPUs"
-                  " (gfx90a:sramecc+:xnack-/gfx942:sramecc+:xnack-/gfx1100)")
+  if( bias.has_value() ) {
+    std::cout << std::endl;
+    std::cout << "Attn_bias sizes : " << bias.value().sizes() << std::endl;
+    std::cout << "attn_bias device: " << bias.value().device() << std::endl;
+	std::cout << "last dim stride: " << bias.value().stride(-1) << std::endl;
   }
 
-  // AOTriton may accept aligned on logsumexp tensor in the future for better
-  // performance, but for now it requires compact logsumexp tensor, even if
-  // compute_logsumexp is false
-  constexpr int kAlignLSE = 1;
+  // Need this in both aot and CK case
+  const auto softmax_scale = sdp::calculate_scale(query, scale).expect_float();
   res = at::empty({B, M, num_heads, Kv}, query.options());
-  logsumexp = at::empty(
+
+  std::cout << "CK Enabled?: " << at::globalContext().getROCmFAPreferredBackend() << std::endl;
+  if(at::globalContext().getROCmFAPreferredBackend() ==
+    at::ROCmFABackend::Ck) {
+    //forward_attention_ck(...);
+    std::cout << "In my branch" << std::endl;
+    std::optional<Tensor> out(res);
+    std::optional<Tensor> seqused_k = std::nullopt;
+    std::optional<Tensor> alibi_slopes = std::nullopt;
+    std::cout << "out(res) dtype " << out.value().dtype();
+
+    auto
+        [out_,
+         q,
+         k,
+         v,
+         lse,
+         seed_t,
+         offset_t,
+         p] =
+            pytorch_flash::mem_eff_forward_ck(
+                                    query,
+                                    key,
+                                    value,
+                                    dropout_p,
+                                    false, // return dropout_randval
+                                    custom_mask_type == 0 ? false : true, // is_causal
+                                    softmax_scale,
+                                    bias,
+                                    out,
+                                    std::nullopt, // cu_seqlens_q: sending in nothing since CKFA works this way
+                                    std::nullopt, // cu_seqlens_k
+                                    seqstart_q,
+                                    seqstart_k,
+                                    std::nullopt,// not passing in optional gen_
+                                    seqused_k);// not passing in optional seqused_k_
+
+    logsumexp = lse;
+  } else { // use aotriton
+    auto ret = aotriton::v2::flash::check_gpu(stream);
+    if (hipSuccess != ret) {
+        TORCH_CHECK(false,
+                  "[AOTriton] Accelerated SDPA only supports MI200/MI300X/Navi31 GPUs"
+                  " (gfx90a:sramecc+:xnack-/gfx942:sramecc+:xnack-/gfx1100)")
+    }
+
+    // AOTriton may accept aligned on logsumexp tensor in the future for better
+    // performance, but for now it requires compact logsumexp tensor, even if
+    // compute_logsumexp is false
+    constexpr int kAlignLSE = 1;
+    logsumexp = at::empty(
       { B, num_heads, max_seqlen_q },
       query.options().dtype(at::ScalarType::Float));
-  at::Tensor softmax_lse = logsumexp.view({B * num_heads, max_seqlen_q});
-  at::Tensor q_t = query.transpose(1, 2);
-  at::Tensor k_t = key.transpose(1, 2);
-  at::Tensor v_t = value.transpose(1, 2);
-  at::Tensor output_t = res.transpose(1, 2);
-  bool is_causal;
-  if (static_cast<int64_t>(sdp::CustomMaskType::CausalFromTopLeft) == custom_mask_type) {
-    is_causal = true;
-  } else if (static_cast<int64_t>(sdp::CustomMaskType::NoCustomMask) == custom_mask_type) {
-    is_causal = false;
-  } else {
-    TORCH_CHECK(false, "[_efficient_attention_forward] Unsupported mask type on ROCM, for now");
-  }
+    at::Tensor softmax_lse = logsumexp.view({B * num_heads, max_seqlen_q});
+    at::Tensor q_t = query.transpose(1, 2);
+    at::Tensor k_t = key.transpose(1, 2);
+    at::Tensor v_t = value.transpose(1, 2);
+    at::Tensor output_t = res.transpose(1, 2);
+    bool is_causal;
+    if (static_cast<int64_t>(sdp::CustomMaskType::CausalFromTopLeft) == custom_mask_type) {
+      is_causal = true;
+    } else if (static_cast<int64_t>(sdp::CustomMaskType::NoCustomMask) == custom_mask_type) {
+      is_causal = false;
+    } else {
+      TORCH_CHECK(false, "[_efficient_attention_forward] Unsupported mask type on ROCM, for now");
+    }
 
-  const auto softmax_scale = sdp::calculate_scale(query, scale).expect_float();
 
-  using aotriton::v2::flash::attn_fwd;
-  using aotriton::v2::flash::attn_fwd_compact_varlen;
-  using sdp::aotriton_adapter::mk_aotensor;
-  using sdp::aotriton_adapter::mk_aoscalartensor;
-  using sdp::aotriton_adapter::mk_philoxtensor;
-  aotriton::TensorView<4> empty_t4(0, {0, 0, 0, 0}, {0, 0, 0, 0}, aotriton::DType::kFloat16);
-  at::Tensor softmax_fa_t = at::empty({ 0, 0, 0, 0 }, query.options());
-  const bool use_philox_state = in_capture_stream;
-  auto seed = use_philox_state ? mk_philoxtensor(philox_state.seed_.ptr) : mk_aoscalartensor(seed_t);
-  auto offset1 = use_philox_state ? mk_philoxtensor(philox_state.offset_.ptr) : mk_aoscalartensor(offset_t);
-  auto offset2 = use_philox_state ? philox_state.offset_intragraph_ : 0;
-  auto seed_output = use_philox_state ? mk_philoxtensor(seed_t.data_ptr<int64_t>()) : mk_philoxtensor(nullptr);
-  auto offset_output = use_philox_state ? mk_philoxtensor(offset_t.data_ptr<int64_t>()) : mk_philoxtensor(nullptr);
-  hipError_t err; // TODO: Error handling
-  if (seqstart_q.has_value()) {
-    // varlen aka nested tensor
-    err = attn_fwd_compact_varlen(mk_aotensor(q_t, "q"),
-                                  mk_aotensor(k_t, "k"),
-                                  mk_aotensor(v_t, "v"),
-                                  mk_aotensor<1>(seqstart_q.value(), "cu_seqlens_q"),
-                                  mk_aotensor<1>(seqstart_k.value(), "cu_seqlens_k"),
-                                  max_seqlen_q,
-                                  max_seqlen_k,
-                                  bias.has_value() ? mk_aotensor(bias.value(), "bias"): empty_t4,
-                                  softmax_scale,
-                                  mk_aotensor<2>(softmax_lse, "M"),
-                                  mk_aotensor(output_t, "Out"),
-                                  dropout_p,
-                                  seed,
-                                  offset1,
-                                  offset2,
-                                  seed_output,
-                                  offset_output,
-                                  mk_aotensor(softmax_fa_t, "encoded_softmax"),
-                                  is_causal,
-                                  stream);
-  } else {
-    err = attn_fwd(mk_aotensor(q_t, "q"),
-                   mk_aotensor(k_t, "k"),
-                   mk_aotensor(v_t, "v"),
-                   bias.has_value() ? mk_aotensor(bias.value(), "bias"): empty_t4,
-                   softmax_scale,
-                   mk_aotensor<2>(softmax_lse, "M"),
-                   mk_aotensor(output_t, "Out"),
-                   dropout_p,
-                   seed,
-                   offset1,
-                   offset2,
-                   seed_output,
-                   offset_output,
-                   mk_aotensor(softmax_fa_t, "encoded_softmax"),
-                   is_causal,
-                   stream);
-  }
-  if (!compute_logsumexp) {
-    // Set the tensor to empty when compute_logsumexp is false
-    logsumexp = at::empty(
+    using aotriton::v2::flash::attn_fwd;
+    using aotriton::v2::flash::attn_fwd_compact_varlen;
+    using sdp::aotriton_adapter::mk_aotensor;
+    using sdp::aotriton_adapter::mk_aoscalartensor;
+    using sdp::aotriton_adapter::mk_philoxtensor;
+    aotriton::TensorView<4> empty_t4(0, {0, 0, 0, 0}, {0, 0, 0, 0}, aotriton::DType::kFloat16);
+    at::Tensor softmax_fa_t = at::empty({ 0, 0, 0, 0 }, query.options());
+    const bool use_philox_state = in_capture_stream;
+    auto seed = use_philox_state ? mk_philoxtensor(philox_state.seed_.ptr) : mk_aoscalartensor(seed_t);
+    auto offset1 = use_philox_state ? mk_philoxtensor(philox_state.offset_.ptr) : mk_aoscalartensor(offset_t);
+    auto offset2 = use_philox_state ? philox_state.offset_intragraph_ : 0;
+    auto seed_output = use_philox_state ? mk_philoxtensor(seed_t.data_ptr<int64_t>()) : mk_philoxtensor(nullptr);
+    auto offset_output = use_philox_state ? mk_philoxtensor(offset_t.data_ptr<int64_t>()) : mk_philoxtensor(nullptr);
+    hipError_t err; // TODO: Error handling
+    if (seqstart_q.has_value()) {
+      // varlen aka nested tensor
+      err = attn_fwd_compact_varlen(mk_aotensor(q_t, "q"),
+                                    mk_aotensor(k_t, "k"),
+                                    mk_aotensor(v_t, "v"),
+                                    mk_aotensor<1>(seqstart_q.value(), "cu_seqlens_q"),
+                                    mk_aotensor<1>(seqstart_k.value(), "cu_seqlens_k"),
+                                    max_seqlen_q,
+                                    max_seqlen_k,
+                                    bias.has_value() ? mk_aotensor(bias.value(), "bias"): empty_t4,
+                                    softmax_scale,
+                                    mk_aotensor<2>(softmax_lse, "M"),
+                                    mk_aotensor(output_t, "Out"),
+                                    dropout_p,
+                                    seed,
+                                    offset1,
+                                    offset2,
+                                    seed_output,
+                                    offset_output,
+                                    mk_aotensor(softmax_fa_t, "encoded_softmax"),
+                                    is_causal,
+                                    stream);
+    } else {
+      err = attn_fwd(mk_aotensor(q_t, "q"),
+                     mk_aotensor(k_t, "k"),
+                     mk_aotensor(v_t, "v"),
+                     bias.has_value() ? mk_aotensor(bias.value(), "bias"): empty_t4,
+                     softmax_scale,
+                     mk_aotensor<2>(softmax_lse, "M"),
+                     mk_aotensor(output_t, "Out"),
+                     dropout_p,
+                     seed,
+                     offset1,
+                     offset2,
+                     seed_output,
+                     offset_output,
+                     mk_aotensor(softmax_fa_t, "encoded_softmax"),
+                     is_causal,
+                     stream);
+    }
+    if (!compute_logsumexp) {
+      // Set the tensor to empty when compute_logsumexp is false
+      logsumexp = at::empty(
         { B * num_heads, max_seqlen_q, 0 },
         query.options().dtype(at::ScalarType::Float));
-  }
+    }
+  } // CK BACKEND
 #else
   // CUDA Implementation
   cudaDeviceProp* p = at::cuda::getDeviceProperties(query.device().index());
@@ -1391,6 +1462,7 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, c10::SymInt, c10::SymInt> _efficient_
   AT_CUDA_CHECK(cudaGetLastError());
 
 #endif // USE_ROCM
+  std::cout << "res dtype: " << res.dtype() << std::endl;
   return std::make_tuple(
       std::move(res),
       std::move(logsumexp),
diff --git a/aten/src/ATen/native/transformers/cuda/attention_backward.cu b/aten/src/ATen/native/transformers/cuda/attention_backward.cu
index 09799ff125d1d..299ae9b15a199 100644
--- a/aten/src/ATen/native/transformers/cuda/attention_backward.cu
+++ b/aten/src/ATen/native/transformers/cuda/attention_backward.cu
@@ -47,6 +47,7 @@
 #include <ATen/native/transformers/hip/aotriton_adapter.h>
 #include <aotriton/flash.h>
 #include <aotriton/runtime.h>
+#include <ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h>
 #endif
 #endif
 
@@ -409,90 +410,133 @@ _efficient_attention_backward(
 
 #ifdef USE_ROCM
   // ROCM Implementation
-  TORCH_CHECK(!num_splits_key.has_value(),
-              "ROCM does not support num_split_keys in _efficient_attention_forward");
-  TORCH_CHECK(!window_size.has_value(),
-              "ROCM does not support window_size in _efficient_attention_forward");
-  auto ret = aotriton::v2::flash::check_gpu(stream);
-  if (hipSuccess != ret) {
-    TORCH_CHECK(false,
+  std::cout << "HITTING CORRECT PATH" << std::endl;
+  std::cout << "bias_requires_grad: " << bias_requires_grad << std::endl;
+  if(at::globalContext().getROCmFAPreferredBackend() == at::ROCmFABackend::Ck)
+  {
+    std::cout << "BACKWARD CK ATTENTION" << std::endl;
+    const auto my_softmax_scale = sdp::calculate_scale(query, scale).expect_float();
+    // Store grad_bias in optional
+    std::optional<at::Tensor> opt_grad_bias = grad_bias;
+    // TODO_ANDY: make sure we are returning the same tensor that is in grad_X
+    auto
+        [dQ,
+         dK,
+         dV,
+         dBias] =
+             pytorch_flash::mem_eff_backward_ck(
+                     grad_out,
+                     query,
+                     key,
+                     value,
+                     out,
+                     logsumexp,
+                     grad_q,
+                     grad_k,
+                     grad_v,
+                     bias,
+                     bias_requires_grad,
+                     opt_grad_bias,
+                     cu_seqlens_q,
+                     cu_seqlens_k,
+                     max_seqlen_q,
+                     max_seqlen_k,
+                     float(dropout_p),
+                     my_softmax_scale,
+                     custom_mask_type == 0 ? false : true, // is_causal
+                     false, // deterministic
+                     false, // zero_tensors
+                     philox_seed,
+                     philox_offset);
+    grad_bias = dBias;
+
+  } else {
+    // Use aotriton
+    TORCH_CHECK(!num_splits_key.has_value(),
+                "ROCM does not support num_split_keys in _efficient_attention_forward");
+    TORCH_CHECK(!window_size.has_value(),
+                "ROCM does not support window_size in _efficient_attention_forward");
+    auto ret = aotriton::v2::flash::check_gpu(stream);
+    if (hipSuccess != ret) {
+      TORCH_CHECK(false,
                 "[AOTriton] Accelerated SDPA only supports MI200/MI300X/Navi31 GPUs"
                 " (gfx90a:sramecc+:xnack-/gfx942:sramecc+:xnack-/gfx1100)")
+    }
+    const auto softmax_scale = sdp::calculate_scale(query, scale).expect_float();
+    bool is_causal;
+    if (static_cast<int64_t>(sdp::CustomMaskType::CausalFromTopLeft) == custom_mask_type) {
+      is_causal = true;
+    } else if (static_cast<int64_t>(sdp::CustomMaskType::NoCustomMask) == custom_mask_type) {
+      is_causal = false;
+    } else {
+      TORCH_CHECK(false, "[_efficient_attention_backward] Unsupported mask type in AOTriton, for now");
+    }
+    at::Tensor q_t = query.permute({0,2,1,3});
+    at::Tensor k_t = key.permute({0,2,1,3});
+    at::Tensor v_t = value.permute({0,2,1,3});
+    at::Tensor out_t = out.permute({0,2,1,3});
+    at::Tensor dq_t = grad_q.permute({0,2,1,3});
+    at::Tensor dk_t = grad_k.permute({0,2,1,3});
+    at::Tensor dv_t = grad_v.permute({0,2,1,3});
+    at::Tensor dout_t = grad_out.permute({0,2,1,3});
+    at::Tensor softmax_lse = logsumexp.view({B * nH, max_seqlen_q});
+    at::Tensor delta = at::empty_like(softmax_lse).contiguous();
+
+    hipError_t err;
+    using aotriton::v2::flash::attn_bwd;
+    using aotriton::v2::flash::attn_bwd_compact_varlen;
+    using sdp::aotriton_adapter::mk_aotensor;
+    using sdp::aotriton_adapter::mk_aoscalartensor;
+    using sdp::aotriton_adapter::cast_dtype;
+    aotriton::TensorView<4> empty_t4(0, {0, 0, 0, 0}, {0, 0, 0, 0}, cast_dtype(query.dtype()));
+    if (cu_seqlens_q.has_value()) {
+      // varlen aka Nested tensor
+      err = attn_bwd_compact_varlen(mk_aotensor(q_t, "q"),
+                                    mk_aotensor(k_t, "k"),
+                                    mk_aotensor(v_t, "v"),
+                                    mk_aotensor<1>(cu_seqlens_q.value(), "cu_seqlens_q"),
+                                    mk_aotensor<1>(cu_seqlens_k.value(), "cu_seqlens_k"),
+                                    max_seqlen_q,
+                                    max_seqlen_k,
+                                    bias.has_value() ? mk_aotensor(bias.value(), "bias") : empty_t4,
+                                    softmax_scale,
+                                    mk_aotensor(out_t, "out"),
+                                    mk_aotensor(dout_t, "dout"),
+                                    mk_aotensor(dq_t, "dq"),
+                                    mk_aotensor(dk_t, "dk"),
+                                    mk_aotensor(dv_t, "dv"),
+                                    bias_requires_grad ? mk_aotensor(grad_bias, "db") : empty_t4,
+                                    mk_aotensor<2>(softmax_lse, "L"),
+                                    mk_aotensor<2>(delta, "delta"),
+                                    float(dropout_p),
+                                    mk_aoscalartensor(philox_seed),
+                                    mk_aoscalartensor(philox_offset),
+                                    0,
+                                    is_causal,
+                                    stream);
+    } else {
+      err = attn_bwd(mk_aotensor(q_t, "q"),
+                     mk_aotensor(k_t, "k"),
+                     mk_aotensor(v_t, "v"),
+                     bias.has_value() ? mk_aotensor(bias.value(), "bias") : empty_t4,
+                     softmax_scale,
+                     mk_aotensor(out_t, "out"),
+                     mk_aotensor(dout_t, "dout"),
+                     mk_aotensor(dq_t, "dq"),
+                     mk_aotensor(dk_t, "dk"),
+                     mk_aotensor(dv_t, "dv"),
+                     bias_requires_grad ? mk_aotensor(grad_bias, "db") : empty_t4,
+                     mk_aotensor<2>(softmax_lse, "L"),
+                     mk_aotensor<2>(delta, "delta"),
+                     float(dropout_p),
+                     mk_aoscalartensor(philox_seed),
+                     mk_aoscalartensor(philox_offset),
+                     0,
+                     is_causal,
+                     stream);
+    }
   }
-  const auto softmax_scale = sdp::calculate_scale(query, scale).expect_float();
-  bool is_causal;
-  if (static_cast<int64_t>(sdp::CustomMaskType::CausalFromTopLeft) == custom_mask_type) {
-    is_causal = true;
-  } else if (static_cast<int64_t>(sdp::CustomMaskType::NoCustomMask) == custom_mask_type) {
-    is_causal = false;
-  } else {
-    TORCH_CHECK(false, "[_efficient_attention_backward] Unsupported mask type in AOTriton, for now");
-  }
-  at::Tensor q_t = query.permute({0,2,1,3});
-  at::Tensor k_t = key.permute({0,2,1,3});
-  at::Tensor v_t = value.permute({0,2,1,3});
-  at::Tensor out_t = out.permute({0,2,1,3});
-  at::Tensor dq_t = grad_q.permute({0,2,1,3});
-  at::Tensor dk_t = grad_k.permute({0,2,1,3});
-  at::Tensor dv_t = grad_v.permute({0,2,1,3});
-  at::Tensor dout_t = grad_out.permute({0,2,1,3});
-  at::Tensor softmax_lse = logsumexp.view({B * nH, max_seqlen_q});
-  at::Tensor delta = at::empty_like(softmax_lse).contiguous();
-
-  hipError_t err;
-  using aotriton::v2::flash::attn_bwd;
-  using aotriton::v2::flash::attn_bwd_compact_varlen;
-  using sdp::aotriton_adapter::mk_aotensor;
-  using sdp::aotriton_adapter::mk_aoscalartensor;
-  using sdp::aotriton_adapter::cast_dtype;
-  aotriton::TensorView<4> empty_t4(0, {0, 0, 0, 0}, {0, 0, 0, 0}, cast_dtype(query.dtype()));
-  if (cu_seqlens_q.has_value()) {
-    // varlen aka Nested tensor
-    err = attn_bwd_compact_varlen(mk_aotensor(q_t, "q"),
-                                  mk_aotensor(k_t, "k"),
-                                  mk_aotensor(v_t, "v"),
-                                  mk_aotensor<1>(cu_seqlens_q.value(), "cu_seqlens_q"),
-                                  mk_aotensor<1>(cu_seqlens_k.value(), "cu_seqlens_k"),
-                                  max_seqlen_q,
-                                  max_seqlen_k,
-                                  bias.has_value() ? mk_aotensor(bias.value(), "bias") : empty_t4,
-                                  softmax_scale,
-                                  mk_aotensor(out_t, "out"),
-                                  mk_aotensor(dout_t, "dout"),
-                                  mk_aotensor(dq_t, "dq"),
-                                  mk_aotensor(dk_t, "dk"),
-                                  mk_aotensor(dv_t, "dv"),
-                                  bias_requires_grad ? mk_aotensor(grad_bias, "db") : empty_t4,
-                                  mk_aotensor<2>(softmax_lse, "L"),
-                                  mk_aotensor<2>(delta, "delta"),
-                                  float(dropout_p),
-                                  mk_aoscalartensor(philox_seed),
-                                  mk_aoscalartensor(philox_offset),
-                                  0,
-                                  is_causal,
-                                  stream);
-  } else {
-    err = attn_bwd(mk_aotensor(q_t, "q"),
-                   mk_aotensor(k_t, "k"),
-                   mk_aotensor(v_t, "v"),
-                   bias.has_value() ? mk_aotensor(bias.value(), "bias") : empty_t4,
-                   softmax_scale,
-                   mk_aotensor(out_t, "out"),
-                   mk_aotensor(dout_t, "dout"),
-                   mk_aotensor(dq_t, "dq"),
-                   mk_aotensor(dk_t, "dk"),
-                   mk_aotensor(dv_t, "dv"),
-                   bias_requires_grad ? mk_aotensor(grad_bias, "db") : empty_t4,
-                   mk_aotensor<2>(softmax_lse, "L"),
-                   mk_aotensor<2>(delta, "delta"),
-                   float(dropout_p),
-                   mk_aoscalartensor(philox_seed),
-                   mk_aoscalartensor(philox_offset),
-                   0,
-                   is_causal,
-                   stream);
-  }
-#else
+#else // USE_CUDA
   at::Tensor workspace;
   cudaDeviceProp* p = at::cuda::getDeviceProperties(query.device().index());
   const int computeCapability = p->major * 10 + p->minor;
@@ -751,6 +795,10 @@ _efficient_attention_backward(
   TORCH_CHECK(kernel_launched, "cutlassB: no kernel found to launch!");
   AT_CUDA_CHECK(cudaGetLastError());
 #endif // USE_ROCM
+  std::cout << "DEVICE_grad_Q: " << grad_q.device() << std::endl;
+  std::cout << "DEVICE_grad_K: " << grad_k.device() << std::endl;
+  std::cout << "DEVICE_grad_V: " << grad_v.device() << std::endl;
+  std::cout << "DEVICE_grad_B: " << grad_bias.device() << std::endl;
   return std::make_tuple(std::move(grad_q), std::move(grad_k), std::move(grad_v), std::move(grad_bias));
   #endif // defined(USE_MEM_EFF_ATTENTION)
   TORCH_CHECK(false, "USE_MEM_EFF_ATTENTION was not enabled for build.")
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/aot/mha_all_aot.hip b/aten/src/ATen/native/transformers/hip/flash_attn/aot/mha_all_aot.hip
index 65425f9c960d2..12ed3e6e1f364 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/aot/mha_all_aot.hip
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/aot/mha_all_aot.hip
@@ -742,4 +742,4 @@ mha_varlen_bwd_aot(const at::Tensor &dout,  // total_q x num_heads, x head_size
 }
 } // namespace pytorch_flash
 
-#endif
+#endif  // USE_FLASH_ATTENTION
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/CMakeLists.txt b/aten/src/ATen/native/transformers/hip/flash_attn/ck/CMakeLists.txt
new file mode 100644
index 0000000000000..a72911cd510eb
--- /dev/null
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/CMakeLists.txt
@@ -0,0 +1,63 @@
+# generate a list of kernels, but not actually emit files at config stage
+execute_process(
+  COMMAND python3 ${CMAKE_CURRENT_LIST_DIR}/../../../../../../../../third_party/composable_kernel/example/ck_tile/01_fmha/generate.py
+  --api fwd --receipt 4 --list_blobs ${CMAKE_CURRENT_LIST_DIR}/fwd_blob_list.txt
+  RESULT_VARIABLE ret
+)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to generate a list of FWD kernels via Python.")
+endif()
+
+execute_process(
+  COMMAND python3 ${CMAKE_CURRENT_LIST_DIR}/../../../../../../../../third_party/composable_kernel/example/ck_tile/01_fmha/generate.py
+  --api bwd --receipt 4 --list_blobs ${CMAKE_CURRENT_LIST_DIR}/bwd_blob_list.txt
+  RESULT_VARIABLE ret
+)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to generate a list of BWD kernels via Python.")
+endif()
+
+# Generate the files for both fwd and bwd
+execute_process(COMMAND python3 ${CMAKE_CURRENT_LIST_DIR}/../../../../../../../../third_party/composable_kernel/example/ck_tile/01_fmha/generate.py --api fwd --receipt 4 --output_dir ${CMAKE_CURRENT_LIST_DIR}
+)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to generate FWD kernels.")
+endif()
+
+execute_process(COMMAND python3 ${CMAKE_CURRENT_LIST_DIR}/../../../../../../../../third_party/composable_kernel/example/ck_tile/01_fmha/generate.py --api bwd --receipt 4 --output_dir ${CMAKE_CURRENT_LIST_DIR}
+  RESULT_VARIABLE ret
+)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to generate BWD kernels.")
+endif()
+
+# Change make_kernel to make_kernel_pt for fwd
+execute_process(
+  COMMAND bash -c "${CMAKE_CURRENT_LIST_DIR}/add_make_kernel_pt.sh ${CMAKE_CURRENT_LIST_DIR}/fwd_blob_list.txt"
+  RESULT_VARIABLE ret)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to change make_kernel to make_kernel_pt for the fwd pass")
+endif()
+
+# Change make_kernel to make_kernel_pt for bwd
+execute_process(
+  COMMAND bash -c "${CMAKE_CURRENT_LIST_DIR}/add_make_kernel_pt.sh ${CMAKE_CURRENT_LIST_DIR}/bwd_blob_list.txt"
+  RESULT_VARIABLE ret)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to change make_kernel to make_kernel_pt for the bwd pass")
+endif()
+
+# Change file extensions to .hip
+execute_process(COMMAND bash -c "for file in ${CMAKE_CURRENT_LIST_DIR}/*.cpp; do mv -- \"$file\" \"\${file%.cpp}.hip\"; done"
+  RESULT_VARIABLE ret
+)
+
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to change the generated instances extensions from .cpp to .hpp")
+endif()
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/add_make_kernel_pt.sh b/aten/src/ATen/native/transformers/hip/flash_attn/ck/add_make_kernel_pt.sh
new file mode 100755
index 0000000000000..672bea1437517
--- /dev/null
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/add_make_kernel_pt.sh
@@ -0,0 +1,30 @@
+#!/bin/bash
+
+# Check if the input file is provided
+if [ "$#" -ne 1 ]; then
+    echo "Usage: $0 <file_list.txt>"
+    exit 1
+fi
+
+# Assign the input file to a variable
+file_list=$1
+
+# Check if the file exists
+if [ ! -f "$file_list" ]; then
+    echo "Error: File '$file_list' not found!"
+    exit 1
+fi
+
+# Loop through each line in the file list
+while IFS= read -r file; do
+    # Check if the file exists in the current directory
+    if [ -f "$file" ]; then
+        # Use sed to replace "make_kernel" with "make_kernel_pt" in place
+        sed -i 's/make_kernel/make_kernel_pt/g' "$file"
+        echo "Updated: $file"
+    else
+        echo "Skipping: $file (not found)"
+    fi
+done < "$file_list"
+
+echo "Replacement completed."
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd.hpp b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd.hpp
index 002b99c9cbf5e..38ec2ef20c5cc 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd.hpp
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_bwd.hpp
@@ -15,11 +15,19 @@
 #include <utility>
 #include <variant>
 
+struct FmhaBwdFp16
+{
+};
+
+struct FmhaBwdBf16
+{
+};
+
 template <typename DataType>
 struct FmhaBwdTypeConfig;
 
 template <>
-struct FmhaBwdTypeConfig<ck_tile::half_t>
+struct FmhaBwdTypeConfig<FmhaBwdFp16>
 {
     using QDataType             = ck_tile::half_t;
     using KDataType             = ck_tile::half_t;
@@ -39,7 +47,7 @@ struct FmhaBwdTypeConfig<ck_tile::half_t>
 };
 
 template <>
-struct FmhaBwdTypeConfig<ck_tile::bf16_t>
+struct FmhaBwdTypeConfig<FmhaBwdBf16>
 {
     using QDataType             = ck_tile::bf16_t;
     using KDataType             = ck_tile::bf16_t;
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_00042c36bc588e60a7c8a9ba297a8a25d8ac0660.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_00042c36bc588e60a7c8a9ba297a8a25d8ac0660.hip
deleted file mode 100644
index 82ce69ab23eb8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_00042c36bc588e60a7c8a9ba297a8a25d8ac0660.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0029076f83a3dc695a167beda6fe19230a2b114b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0029076f83a3dc695a167beda6fe19230a2b114b.hip
deleted file mode 100644
index b9b428a079f8d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0029076f83a3dc695a167beda6fe19230a2b114b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_006c417a52a1bd7c55e45d111483d26f4480caeb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_006c417a52a1bd7c55e45d111483d26f4480caeb.hip
deleted file mode 100644
index 3ed1294bb7b48..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_006c417a52a1bd7c55e45d111483d26f4480caeb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_008f2429c678d13386a06e8d8b15c4b480940ff3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_008f2429c678d13386a06e8d8b15c4b480940ff3.hip
deleted file mode 100644
index f7c9761d72109..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_008f2429c678d13386a06e8d8b15c4b480940ff3.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_00a2adbe938d458d51ca5fc4020667a215b672a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_00a2adbe938d458d51ca5fc4020667a215b672a4.hip
deleted file mode 100644
index fd3344b077ab0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_00a2adbe938d458d51ca5fc4020667a215b672a4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_012c0f480917c329f4c3c6c666cf32af2d82b294.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_012c0f480917c329f4c3c6c666cf32af2d82b294.hip
deleted file mode 100644
index e74476990d446..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_012c0f480917c329f4c3c6c666cf32af2d82b294.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_014c209d5cfc6b965bfd78c64bf132c0154e32be.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_014c209d5cfc6b965bfd78c64bf132c0154e32be.hip
deleted file mode 100644
index 8ca5e5d1cded0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_014c209d5cfc6b965bfd78c64bf132c0154e32be.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0153ec18d3ded0f8bdc6459ea5757ebd94d9faf2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0153ec18d3ded0f8bdc6459ea5757ebd94d9faf2.hip
deleted file mode 100644
index a95b27e98885a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0153ec18d3ded0f8bdc6459ea5757ebd94d9faf2.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ac1a2ecf9a487809e46faa92e267df2d47de91.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ac1a2ecf9a487809e46faa92e267df2d47de91.hip
deleted file mode 100644
index 54a4aa16302e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ac1a2ecf9a487809e46faa92e267df2d47de91.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ca79005067e20e4eed5a72ff9187cde702cd1c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ca79005067e20e4eed5a72ff9187cde702cd1c.hip
deleted file mode 100644
index 5244d6b6336ac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ca79005067e20e4eed5a72ff9187cde702cd1c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01cb354dddef6e99e4ac843f2adafcddfc58d520.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01cb354dddef6e99e4ac843f2adafcddfc58d520.hip
deleted file mode 100644
index e9cf79f84b3f6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01cb354dddef6e99e4ac843f2adafcddfc58d520.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01d12033d59ce2799a2a024e5d9232325ccf1320.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01d12033d59ce2799a2a024e5d9232325ccf1320.hip
deleted file mode 100644
index 5becf11a95a8f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01d12033d59ce2799a2a024e5d9232325ccf1320.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01d3b034a2d8d0b83c0aefa4faac6c3f28ce737f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01d3b034a2d8d0b83c0aefa4faac6c3f28ce737f.hip
deleted file mode 100644
index 49470c9aed7de..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01d3b034a2d8d0b83c0aefa4faac6c3f28ce737f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e2428c5447aa9a78f79f73f31cf685c586872d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e2428c5447aa9a78f79f73f31cf685c586872d.hip
deleted file mode 100644
index fff9e8ea4f6a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e2428c5447aa9a78f79f73f31cf685c586872d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e8aedb7b7d77f44a46b2e9b7a826f245aaf4a7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e8aedb7b7d77f44a46b2e9b7a826f245aaf4a7.hip
deleted file mode 100644
index fdee95116f798..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e8aedb7b7d77f44a46b2e9b7a826f245aaf4a7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e8f0df0c54ce619e5b66441b3c96a5e18b05d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e8f0df0c54ce619e5b66441b3c96a5e18b05d6.hip
deleted file mode 100644
index b2c359694a713..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01e8f0df0c54ce619e5b66441b3c96a5e18b05d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ee0083f6df962c4a754cd3295b1a436c590a0e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ee0083f6df962c4a754cd3295b1a436c590a0e.hip
deleted file mode 100644
index 4dda2f0d57486..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01ee0083f6df962c4a754cd3295b1a436c590a0e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01f74764c3c3284fdd1b67d0ea781c2261ed0de6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01f74764c3c3284fdd1b67d0ea781c2261ed0de6.hip
deleted file mode 100644
index c09ec03911199..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_01f74764c3c3284fdd1b67d0ea781c2261ed0de6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0225857454eaab2eb664aef7a0849ce12c32fdf9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0225857454eaab2eb664aef7a0849ce12c32fdf9.hip
deleted file mode 100644
index 08b088d8cb3ef..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0225857454eaab2eb664aef7a0849ce12c32fdf9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0237c76137df14fb808ade8bd6837045f2aaa5c9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0237c76137df14fb808ade8bd6837045f2aaa5c9.hip
deleted file mode 100644
index 0cb05acfe8bf3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0237c76137df14fb808ade8bd6837045f2aaa5c9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0271bd8b7c270e1593871b638288a4923342c446.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0271bd8b7c270e1593871b638288a4923342c446.hip
deleted file mode 100644
index 325e4cbe881bf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0271bd8b7c270e1593871b638288a4923342c446.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_02d88a03cd3966dd0cff550065f58c3ffecfff6c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_02d88a03cd3966dd0cff550065f58c3ffecfff6c.hip
deleted file mode 100644
index 0e578836ab414..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_02d88a03cd3966dd0cff550065f58c3ffecfff6c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_02ff94e3c787a7b06ffc90c25777fa74f225e32c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_02ff94e3c787a7b06ffc90c25777fa74f225e32c.hip
deleted file mode 100644
index e7cd9d9065849..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_02ff94e3c787a7b06ffc90c25777fa74f225e32c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_030a759dcc92028b4c6f317fc230b98cb929e806.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_030a759dcc92028b4c6f317fc230b98cb929e806.hip
deleted file mode 100644
index ae0bf3c79c8a5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_030a759dcc92028b4c6f317fc230b98cb929e806.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_031b12f9fd94e01aaff2c0da4f35f346822087e4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_031b12f9fd94e01aaff2c0da4f35f346822087e4.hip
deleted file mode 100644
index 996baf438ef4c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_031b12f9fd94e01aaff2c0da4f35f346822087e4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_036887daf6cc092e7422a17882488e59cecfb643.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_036887daf6cc092e7422a17882488e59cecfb643.hip
deleted file mode 100644
index 6287aac677e22..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_036887daf6cc092e7422a17882488e59cecfb643.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_037c6c80fcec3eb8b0bef50ad6af6d27bf5447f5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_037c6c80fcec3eb8b0bef50ad6af6d27bf5447f5.hip
deleted file mode 100644
index 91c2ab28c29f6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_037c6c80fcec3eb8b0bef50ad6af6d27bf5447f5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0392491c5a6dfc742c2be483419a40f6a7a7ea56.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0392491c5a6dfc742c2be483419a40f6a7a7ea56.hip
deleted file mode 100644
index 4052163433113..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0392491c5a6dfc742c2be483419a40f6a7a7ea56.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_03a71615a088e972c998f9c7cb44566c268c5124.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_03a71615a088e972c998f9c7cb44566c268c5124.hip
deleted file mode 100644
index ef626eecd7daf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_03a71615a088e972c998f9c7cb44566c268c5124.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_03ff035717140f7385282419598cb4fb2881ce8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_03ff035717140f7385282419598cb4fb2881ce8e.hip
deleted file mode 100644
index 805b8dc0a38d2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_03ff035717140f7385282419598cb4fb2881ce8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_041a0718891596ddac1fb0088637029233ccbe60.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_041a0718891596ddac1fb0088637029233ccbe60.hip
deleted file mode 100644
index 3e4a7c904e36d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_041a0718891596ddac1fb0088637029233ccbe60.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_042a156e9eb935555ab14a84461959b466c2fb5b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_042a156e9eb935555ab14a84461959b466c2fb5b.hip
deleted file mode 100644
index f6191ab20bc2a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_042a156e9eb935555ab14a84461959b466c2fb5b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04641230fe9a50a221047f7a1df8a370f72805b9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04641230fe9a50a221047f7a1df8a370f72805b9.hip
deleted file mode 100644
index 59ab7b1f25848..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04641230fe9a50a221047f7a1df8a370f72805b9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04c363e11d202c6d2f4bb753661c5a2043edc0ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04c363e11d202c6d2f4bb753661c5a2043edc0ad.hip
deleted file mode 100644
index 7269255197c1e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04c363e11d202c6d2f4bb753661c5a2043edc0ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04caeecbc01667ec6f5599358a0a20423aa9a00b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04caeecbc01667ec6f5599358a0a20423aa9a00b.hip
deleted file mode 100644
index b3613e705745d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04caeecbc01667ec6f5599358a0a20423aa9a00b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04f39b453505f68a5091f68b1c3de48369d1e7ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04f39b453505f68a5091f68b1c3de48369d1e7ea.hip
deleted file mode 100644
index 3518c4f0ee3b0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04f39b453505f68a5091f68b1c3de48369d1e7ea.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04ffca078cfab8bc6c4ccd1cc8994a1bb4a88ea7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04ffca078cfab8bc6c4ccd1cc8994a1bb4a88ea7.hip
deleted file mode 100644
index 8d967cb643a69..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_04ffca078cfab8bc6c4ccd1cc8994a1bb4a88ea7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0502e718337eab7d47aa65cea7d3c5f641484520.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0502e718337eab7d47aa65cea7d3c5f641484520.hip
deleted file mode 100644
index 8054f38064ee7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0502e718337eab7d47aa65cea7d3c5f641484520.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0513b2f3bd8ad51315aadb7f63737201898adca8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0513b2f3bd8ad51315aadb7f63737201898adca8.hip
deleted file mode 100644
index 7e5bc574a2fd9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0513b2f3bd8ad51315aadb7f63737201898adca8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_053981d9e7af2ebc0f91e61ac5e25cbe68c95bd8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_053981d9e7af2ebc0f91e61ac5e25cbe68c95bd8.hip
deleted file mode 100644
index f475e4b4883e4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_053981d9e7af2ebc0f91e61ac5e25cbe68c95bd8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_054fda16133a0d25077967b05425f9128e1fe1a5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_054fda16133a0d25077967b05425f9128e1fe1a5.hip
deleted file mode 100644
index 777a1c0f9e951..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_054fda16133a0d25077967b05425f9128e1fe1a5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05538339c21c92c53d237865d72debaaf2ee5075.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05538339c21c92c53d237865d72debaaf2ee5075.hip
deleted file mode 100644
index 6f7e82be7e1ef..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05538339c21c92c53d237865d72debaaf2ee5075.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0595316f0dfffda03e5296b959a49ec3f3c48d67.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0595316f0dfffda03e5296b959a49ec3f3c48d67.hip
deleted file mode 100644
index f46f2b4f372bd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0595316f0dfffda03e5296b959a49ec3f3c48d67.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05dfe927fd64a564c5fad537fb7c41ee9c94c2c0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05dfe927fd64a564c5fad537fb7c41ee9c94c2c0.hip
deleted file mode 100644
index 486f54cbef9ce..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05dfe927fd64a564c5fad537fb7c41ee9c94c2c0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05e60b3ab7477f9edc8576a8bf43e3a62b8d5ef8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05e60b3ab7477f9edc8576a8bf43e3a62b8d5ef8.hip
deleted file mode 100644
index 90659871cb42c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05e60b3ab7477f9edc8576a8bf43e3a62b8d5ef8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05f794c7023cbb7e35f1fd1ae45bd2377bfbc520.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05f794c7023cbb7e35f1fd1ae45bd2377bfbc520.hip
deleted file mode 100644
index da8fdbd723df0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_05f794c7023cbb7e35f1fd1ae45bd2377bfbc520.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0628931bf5cc1daa6e106cf60bb21fa1aac6b1df.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0628931bf5cc1daa6e106cf60bb21fa1aac6b1df.hip
deleted file mode 100644
index 85d85fe6f34eb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0628931bf5cc1daa6e106cf60bb21fa1aac6b1df.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_062c8c3c1cf6c33af4574099e9b6ac54a55ad776.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_062c8c3c1cf6c33af4574099e9b6ac54a55ad776.hip
deleted file mode 100644
index c1bac538b6160..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_062c8c3c1cf6c33af4574099e9b6ac54a55ad776.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0682150e93f547e00f13cd8984779bf49b91e50c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0682150e93f547e00f13cd8984779bf49b91e50c.hip
deleted file mode 100644
index 3ac2e48e176de..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0682150e93f547e00f13cd8984779bf49b91e50c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_069c663be0267c009be4814e9e4e7c13ec999411.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_069c663be0267c009be4814e9e4e7c13ec999411.hip
deleted file mode 100644
index 51937df035c80..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_069c663be0267c009be4814e9e4e7c13ec999411.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06ae52ef937cc27c544e32025ea0dadb7fad982d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06ae52ef937cc27c544e32025ea0dadb7fad982d.hip
deleted file mode 100644
index 04a4bd9026f3b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06ae52ef937cc27c544e32025ea0dadb7fad982d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06b74acd9abfbd1c4ec2f4c718eeb92a0bca7bab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06b74acd9abfbd1c4ec2f4c718eeb92a0bca7bab.hip
deleted file mode 100644
index ce8d490113523..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06b74acd9abfbd1c4ec2f4c718eeb92a0bca7bab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06ba94794a14f0f0022af6f5f3c16e1e16959d4c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06ba94794a14f0f0022af6f5f3c16e1e16959d4c.hip
deleted file mode 100644
index cd324663db2ee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_06ba94794a14f0f0022af6f5f3c16e1e16959d4c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_071751b1012b90f7b57f8591cd06ae1fd27d9cd3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_071751b1012b90f7b57f8591cd06ae1fd27d9cd3.hip
deleted file mode 100644
index 6f184966dbf7e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_071751b1012b90f7b57f8591cd06ae1fd27d9cd3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0766e7aa4b263a811408b285213e47176ee2bdaf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0766e7aa4b263a811408b285213e47176ee2bdaf.hip
deleted file mode 100644
index 7b86f6cf29554..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0766e7aa4b263a811408b285213e47176ee2bdaf.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_076b3beb57b30afb30636f948e3989b346b38d20.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_076b3beb57b30afb30636f948e3989b346b38d20.hip
deleted file mode 100644
index 7638d922dd73a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_076b3beb57b30afb30636f948e3989b346b38d20.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0789852b0cd3cc030c78b28f2fd5b6b0546382a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0789852b0cd3cc030c78b28f2fd5b6b0546382a4.hip
deleted file mode 100644
index 49714c5132717..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0789852b0cd3cc030c78b28f2fd5b6b0546382a4.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_078b96ad691a85eebd18586db0b62b8911016d9c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_078b96ad691a85eebd18586db0b62b8911016d9c.hip
deleted file mode 100644
index b2668f3d0d3b2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_078b96ad691a85eebd18586db0b62b8911016d9c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_07c3fc96d2bebe546dce6ebf46e5c7a519959599.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_07c3fc96d2bebe546dce6ebf46e5c7a519959599.hip
deleted file mode 100644
index bb4930d972db0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_07c3fc96d2bebe546dce6ebf46e5c7a519959599.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_07ff04fcc273e469737512893ea3fb5876ac131d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_07ff04fcc273e469737512893ea3fb5876ac131d.hip
deleted file mode 100644
index 68d9e794357a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_07ff04fcc273e469737512893ea3fb5876ac131d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0801c56831b4c6428200db6318638a2129bb197a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0801c56831b4c6428200db6318638a2129bb197a.hip
deleted file mode 100644
index e0dd05a50330b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0801c56831b4c6428200db6318638a2129bb197a.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0836d5dfc0f939ab9a4064b403339373caf35b56.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0836d5dfc0f939ab9a4064b403339373caf35b56.hip
deleted file mode 100644
index 188be6dbcaa90..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0836d5dfc0f939ab9a4064b403339373caf35b56.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0842c4e3aabdf55405b3ce09ce1899245ddf11ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0842c4e3aabdf55405b3ce09ce1899245ddf11ad.hip
deleted file mode 100644
index a7e145637b17e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0842c4e3aabdf55405b3ce09ce1899245ddf11ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_085722b43cde5f37242edb071f639da7c4a0bd48.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_085722b43cde5f37242edb071f639da7c4a0bd48.hip
deleted file mode 100644
index 2b1591bea8dc5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_085722b43cde5f37242edb071f639da7c4a0bd48.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0878b9aa31429d23a93cd953cc6a2fc5f43d0d3a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0878b9aa31429d23a93cd953cc6a2fc5f43d0d3a.hip
deleted file mode 100644
index 85b9289f71321..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0878b9aa31429d23a93cd953cc6a2fc5f43d0d3a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_089a347aef8a920e3b59d5ffe71fc5bfe002609c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_089a347aef8a920e3b59d5ffe71fc5bfe002609c.hip
deleted file mode 100644
index 7c6617552ed57..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_089a347aef8a920e3b59d5ffe71fc5bfe002609c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_089de13222caec1483207d4a54249f8da4f9c151.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_089de13222caec1483207d4a54249f8da4f9c151.hip
deleted file mode 100644
index d214aec72e7d1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_089de13222caec1483207d4a54249f8da4f9c151.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_091cb49c1958fb4342d79f367ea93cf2b472f785.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_091cb49c1958fb4342d79f367ea93cf2b472f785.hip
deleted file mode 100644
index 3e7612dcfb1bb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_091cb49c1958fb4342d79f367ea93cf2b472f785.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_093834d4d3fe76e1745e4482c6b51b550c6f3dfc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_093834d4d3fe76e1745e4482c6b51b550c6f3dfc.hip
deleted file mode 100644
index 1b65db905e3fb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_093834d4d3fe76e1745e4482c6b51b550c6f3dfc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09513bff5c1da6aadf11d2e8272a422eabff21bc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09513bff5c1da6aadf11d2e8272a422eabff21bc.hip
deleted file mode 100644
index d932430b16126..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09513bff5c1da6aadf11d2e8272a422eabff21bc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_096863cd93d1b105a617d0daa1d4f37d7fb6b893.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_096863cd93d1b105a617d0daa1d4f37d7fb6b893.hip
deleted file mode 100644
index 99cf5eb1411fe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_096863cd93d1b105a617d0daa1d4f37d7fb6b893.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0968cebd81ade762c2f92fffc0153fa7a2b91eb5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0968cebd81ade762c2f92fffc0153fa7a2b91eb5.hip
deleted file mode 100644
index 2df3d42336a33..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0968cebd81ade762c2f92fffc0153fa7a2b91eb5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_096e888c52d0f4a5847d7515fcc66208b1ff40d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_096e888c52d0f4a5847d7515fcc66208b1ff40d3.hip
deleted file mode 100644
index 579b87f23f403..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_096e888c52d0f4a5847d7515fcc66208b1ff40d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_097b3e1dae9bfb2e89398706508f8e01966fd4ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_097b3e1dae9bfb2e89398706508f8e01966fd4ea.hip
deleted file mode 100644
index 0a105968b2d46..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_097b3e1dae9bfb2e89398706508f8e01966fd4ea.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09d76cca48b71dbcc9bd96734787209fee4c9a74.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09d76cca48b71dbcc9bd96734787209fee4c9a74.hip
deleted file mode 100644
index 2382f25695206..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09d76cca48b71dbcc9bd96734787209fee4c9a74.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09e50367b62bb09071e28b44235a7c112645a706.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09e50367b62bb09071e28b44235a7c112645a706.hip
deleted file mode 100644
index c6c9a21f84fd2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09e50367b62bb09071e28b44235a7c112645a706.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09ecb6347009f6a5d5530a6acf90f9f40288cbcf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09ecb6347009f6a5d5530a6acf90f9f40288cbcf.hip
deleted file mode 100644
index fe5ee00e6d678..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_09ecb6347009f6a5d5530a6acf90f9f40288cbcf.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a2b116fd5065109aae46ee547e4f49ad0e9d6e1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a2b116fd5065109aae46ee547e4f49ad0e9d6e1.hip
deleted file mode 100644
index b3b32d1d9e4f9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a2b116fd5065109aae46ee547e4f49ad0e9d6e1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a4e76d89b175e1d9fd2e9fb908d5fce1ebb945d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a4e76d89b175e1d9fd2e9fb908d5fce1ebb945d.hip
deleted file mode 100644
index da2e456243bdc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a4e76d89b175e1d9fd2e9fb908d5fce1ebb945d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a55ed15ef58c941e06dda890aeb530e28eb7bba.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a55ed15ef58c941e06dda890aeb530e28eb7bba.hip
deleted file mode 100644
index 758b29a936950..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a55ed15ef58c941e06dda890aeb530e28eb7bba.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a672fca51de618e3441cf8764e8e83eb782f2c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a672fca51de618e3441cf8764e8e83eb782f2c7.hip
deleted file mode 100644
index c6fe88763aeee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a672fca51de618e3441cf8764e8e83eb782f2c7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a68c2f9a3acdd787b81be455cbc7836c8bfd90c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a68c2f9a3acdd787b81be455cbc7836c8bfd90c.hip
deleted file mode 100644
index e61f57695693d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a68c2f9a3acdd787b81be455cbc7836c8bfd90c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a89417a043556970f72eebd48b4f3e7ac15377a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a89417a043556970f72eebd48b4f3e7ac15377a.hip
deleted file mode 100644
index 6ca2ca060c396..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a89417a043556970f72eebd48b4f3e7ac15377a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a92671b6ea99891c0d69b1c793f4d131b9a82ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a92671b6ea99891c0d69b1c793f4d131b9a82ed.hip
deleted file mode 100644
index 75ee19bea8d46..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0a92671b6ea99891c0d69b1c793f4d131b9a82ed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0aafb881e34a3794970a1282af740b3f19c138b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0aafb881e34a3794970a1282af740b3f19c138b1.hip
deleted file mode 100644
index 6e28dab684c10..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0aafb881e34a3794970a1282af740b3f19c138b1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ace6e29e1d3060c3086c08fe27b471e375f9c75.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ace6e29e1d3060c3086c08fe27b471e375f9c75.hip
deleted file mode 100644
index 55f55e980f2d2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ace6e29e1d3060c3086c08fe27b471e375f9c75.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ad9d68fcee021437e13ffdf94d78252205f5a31.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ad9d68fcee021437e13ffdf94d78252205f5a31.hip
deleted file mode 100644
index fe4433b0d9940..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ad9d68fcee021437e13ffdf94d78252205f5a31.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b2647b5982405a48e8c8888552a4b89386ccdd9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b2647b5982405a48e8c8888552a4b89386ccdd9.hip
deleted file mode 100644
index 04bafdae15d73..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b2647b5982405a48e8c8888552a4b89386ccdd9.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b2efefea81036641561bed80c75d77651176f74.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b2efefea81036641561bed80c75d77651176f74.hip
deleted file mode 100644
index 617f65d6a870a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b2efefea81036641561bed80c75d77651176f74.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b3153af7bcdba33115a0d31f121fd76be2ffbcc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b3153af7bcdba33115a0d31f121fd76be2ffbcc.hip
deleted file mode 100644
index 8859071d07079..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b3153af7bcdba33115a0d31f121fd76be2ffbcc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b532fcf26f90c82a792cde7943634f667c1d033.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b532fcf26f90c82a792cde7943634f667c1d033.hip
deleted file mode 100644
index fbec5b827fee2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b532fcf26f90c82a792cde7943634f667c1d033.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b90a0186d8b8004e3f19886c7992c8e04d0e066.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b90a0186d8b8004e3f19886c7992c8e04d0e066.hip
deleted file mode 100644
index fed24fa0e95ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b90a0186d8b8004e3f19886c7992c8e04d0e066.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b9585ba1c10acf67115c5899b3546608541820d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b9585ba1c10acf67115c5899b3546608541820d.hip
deleted file mode 100644
index d615a130ca32d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0b9585ba1c10acf67115c5899b3546608541820d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0bb81407c8a2b3cdc5fecf655b3ad64d5d729cc9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0bb81407c8a2b3cdc5fecf655b3ad64d5d729cc9.hip
deleted file mode 100644
index 5125929e3beaa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0bb81407c8a2b3cdc5fecf655b3ad64d5d729cc9.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0bc7910aac798f0555e9e505ad7f177c9fbbd92c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0bc7910aac798f0555e9e505ad7f177c9fbbd92c.hip
deleted file mode 100644
index 64520c318a5d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0bc7910aac798f0555e9e505ad7f177c9fbbd92c.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0be8cf70c6be969ecfca675782c860b5b75ac089.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0be8cf70c6be969ecfca675782c860b5b75ac089.hip
deleted file mode 100644
index e6477808a9b90..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0be8cf70c6be969ecfca675782c860b5b75ac089.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0befed50a89d80c22b2c8c3d5ba67d73c3d0190e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0befed50a89d80c22b2c8c3d5ba67d73c3d0190e.hip
deleted file mode 100644
index b6f0661741a89..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0befed50a89d80c22b2c8c3d5ba67d73c3d0190e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c32a2d9701e23dd930119c4ee8089042b5b0ac5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c32a2d9701e23dd930119c4ee8089042b5b0ac5.hip
deleted file mode 100644
index 85b926d7b8bbb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c32a2d9701e23dd930119c4ee8089042b5b0ac5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c3b2ec99fa7b09c7f78dcc3142a661d686044ac.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c3b2ec99fa7b09c7f78dcc3142a661d686044ac.hip
deleted file mode 100644
index bcc0e1185d46b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c3b2ec99fa7b09c7f78dcc3142a661d686044ac.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c8a0bb89a6f05289c0405df5126fa0cc16252e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c8a0bb89a6f05289c0405df5126fa0cc16252e7.hip
deleted file mode 100644
index c573d72c183a1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c8a0bb89a6f05289c0405df5126fa0cc16252e7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c93c65e5942a2f43f2e491547add02777dd2eee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c93c65e5942a2f43f2e491547add02777dd2eee.hip
deleted file mode 100644
index 8849251514549..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c93c65e5942a2f43f2e491547add02777dd2eee.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c9bd38b8f9009d932ec49204fdea39a52885246.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c9bd38b8f9009d932ec49204fdea39a52885246.hip
deleted file mode 100644
index 1126191524705..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0c9bd38b8f9009d932ec49204fdea39a52885246.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0caeedaa7d50f1741d618fb6c573529eebb075b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0caeedaa7d50f1741d618fb6c573529eebb075b1.hip
deleted file mode 100644
index 95c691d24d3c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0caeedaa7d50f1741d618fb6c573529eebb075b1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0cdef49859c80c6b3ba18eb2fb4c35c72abc1cf2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0cdef49859c80c6b3ba18eb2fb4c35c72abc1cf2.hip
deleted file mode 100644
index 331d0079fcd91..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0cdef49859c80c6b3ba18eb2fb4c35c72abc1cf2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0cee6b9427c164d78994150305a47f73954a67c0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0cee6b9427c164d78994150305a47f73954a67c0.hip
deleted file mode 100644
index cd7110665afb5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0cee6b9427c164d78994150305a47f73954a67c0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0d0e0147a92061d32608a34e7b47bd534eb787fa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0d0e0147a92061d32608a34e7b47bd534eb787fa.hip
deleted file mode 100644
index 74aab67e6f826..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0d0e0147a92061d32608a34e7b47bd534eb787fa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0d13a4c8d169877da6408584dc1f20a6f7c5e3aa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0d13a4c8d169877da6408584dc1f20a6f7c5e3aa.hip
deleted file mode 100644
index 43fe4e0fdd16c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0d13a4c8d169877da6408584dc1f20a6f7c5e3aa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0dde401aa76cb5425563cbbdb0362748148da3ca.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0dde401aa76cb5425563cbbdb0362748148da3ca.hip
deleted file mode 100644
index 56ed4317ebe9f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0dde401aa76cb5425563cbbdb0362748148da3ca.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e007c36231ccdae12f102eacca1f74b0711b9c6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e007c36231ccdae12f102eacca1f74b0711b9c6.hip
deleted file mode 100644
index a16ef08dff9f7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e007c36231ccdae12f102eacca1f74b0711b9c6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e0a2370f2a320484d8f9f21e3197425c2dbe9ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e0a2370f2a320484d8f9f21e3197425c2dbe9ad.hip
deleted file mode 100644
index 2de437aaeab2c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e0a2370f2a320484d8f9f21e3197425c2dbe9ad.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e1dbc9c433ce8ec33ace9e62550261d613db582.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e1dbc9c433ce8ec33ace9e62550261d613db582.hip
deleted file mode 100644
index 1961b436f0b36..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e1dbc9c433ce8ec33ace9e62550261d613db582.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e3f4cd28a4c06cc109f6a0798a77844bcc750b7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e3f4cd28a4c06cc109f6a0798a77844bcc750b7.hip
deleted file mode 100644
index 64e8dc795f95e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e3f4cd28a4c06cc109f6a0798a77844bcc750b7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e661b5f30566d1f159f060c264849c7ae4772f1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e661b5f30566d1f159f060c264849c7ae4772f1.hip
deleted file mode 100644
index 321c06e27fcf0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0e661b5f30566d1f159f060c264849c7ae4772f1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ebacd06455ab20eba78b389462946716b5819f6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ebacd06455ab20eba78b389462946716b5819f6.hip
deleted file mode 100644
index 5041498b6447f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ebacd06455ab20eba78b389462946716b5819f6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ef309b923172f4c0fb38d9b9f5325b33b4877c2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ef309b923172f4c0fb38d9b9f5325b33b4877c2.hip
deleted file mode 100644
index a86b86e858c1b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ef309b923172f4c0fb38d9b9f5325b33b4877c2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ef9b9413697d6f4573c6605bff6f58d027c5016.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ef9b9413697d6f4573c6605bff6f58d027c5016.hip
deleted file mode 100644
index d717a3afff85e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0ef9b9413697d6f4573c6605bff6f58d027c5016.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0efdaa9266a5a464009297dc59db92504f8bf1a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0efdaa9266a5a464009297dc59db92504f8bf1a3.hip
deleted file mode 100644
index 8a351a5edd3a9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0efdaa9266a5a464009297dc59db92504f8bf1a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0f0c699d9c3b0ed62097e38ba05e40e815cf474e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0f0c699d9c3b0ed62097e38ba05e40e815cf474e.hip
deleted file mode 100644
index 24ffbe00a1911..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0f0c699d9c3b0ed62097e38ba05e40e815cf474e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0f588dcb2ef86677ebf84e406eb802e9921d1f1e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0f588dcb2ef86677ebf84e406eb802e9921d1f1e.hip
deleted file mode 100644
index 09e177b2d39eb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0f588dcb2ef86677ebf84e406eb802e9921d1f1e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fbb0bef3b388867e75d7a8a187b8b4b650a42ae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fbb0bef3b388867e75d7a8a187b8b4b650a42ae.hip
deleted file mode 100644
index 572d4bbd651b9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fbb0bef3b388867e75d7a8a187b8b4b650a42ae.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fbddf533661642d84bf5a16149692d5a892182a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fbddf533661642d84bf5a16149692d5a892182a.hip
deleted file mode 100644
index ea5280dfceffb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fbddf533661642d84bf5a16149692d5a892182a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fcb7492feb79e27e0bda73e57ef7dab410e2bb6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fcb7492feb79e27e0bda73e57ef7dab410e2bb6.hip
deleted file mode 100644
index b3db703d66ed7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fcb7492feb79e27e0bda73e57ef7dab410e2bb6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fd4068ea93fcf4df463e3bf3a6898d23b65da7f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fd4068ea93fcf4df463e3bf3a6898d23b65da7f.hip
deleted file mode 100644
index 3d721fee05b84..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_0fd4068ea93fcf4df463e3bf3a6898d23b65da7f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_103186dbad604763008e0204a1ea90baecef8877.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_103186dbad604763008e0204a1ea90baecef8877.hip
deleted file mode 100644
index cd6beecfbde11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_103186dbad604763008e0204a1ea90baecef8877.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1037f1bc50c4a65dac09ba56b701256b701c4322.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1037f1bc50c4a65dac09ba56b701256b701c4322.hip
deleted file mode 100644
index dfb8780026c56..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1037f1bc50c4a65dac09ba56b701256b701c4322.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10a055e5c3d6a953d470db5dc21449766248058a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10a055e5c3d6a953d470db5dc21449766248058a.hip
deleted file mode 100644
index 23bf3d8a41aae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10a055e5c3d6a953d470db5dc21449766248058a.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10c24f1f9009e46afa3a59193784cc2575f79056.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10c24f1f9009e46afa3a59193784cc2575f79056.hip
deleted file mode 100644
index f12904567d57a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10c24f1f9009e46afa3a59193784cc2575f79056.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10ceed95b0a0a01f844678717c88e0426fb503fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10ceed95b0a0a01f844678717c88e0426fb503fd.hip
deleted file mode 100644
index 05cfb54d3ee45..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_10ceed95b0a0a01f844678717c88e0426fb503fd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1132b11429034d96d82c82dbfdb69e460ad8a564.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1132b11429034d96d82c82dbfdb69e460ad8a564.hip
deleted file mode 100644
index 7f9f7bb8ccde0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1132b11429034d96d82c82dbfdb69e460ad8a564.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_11e7df31541c3aa919e9825ad7dc4432f9a03c0c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_11e7df31541c3aa919e9825ad7dc4432f9a03c0c.hip
deleted file mode 100644
index f40c23bce7e72..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_11e7df31541c3aa919e9825ad7dc4432f9a03c0c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_11ff174ff2175e9ec22ac3a0fa59dd7713b79643.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_11ff174ff2175e9ec22ac3a0fa59dd7713b79643.hip
deleted file mode 100644
index 29a1861d8aad8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_11ff174ff2175e9ec22ac3a0fa59dd7713b79643.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1211733062ed30b876f1d63bffa642d77e258dd6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1211733062ed30b876f1d63bffa642d77e258dd6.hip
deleted file mode 100644
index 3716bde6947d0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1211733062ed30b876f1d63bffa642d77e258dd6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12207f4b6e7fac27d6c16493a5373f448a2aaae8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12207f4b6e7fac27d6c16493a5373f448a2aaae8.hip
deleted file mode 100644
index e5220100370e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12207f4b6e7fac27d6c16493a5373f448a2aaae8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1241814f76107d74ed069ecec99a248676487eee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1241814f76107d74ed069ecec99a248676487eee.hip
deleted file mode 100644
index dd665cc98c952..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1241814f76107d74ed069ecec99a248676487eee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12d5c8a4988efe60ef7943ecd73e18a28a736583.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12d5c8a4988efe60ef7943ecd73e18a28a736583.hip
deleted file mode 100644
index 4607dd0a0afdb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12d5c8a4988efe60ef7943ecd73e18a28a736583.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12d60c8abecb3bc9b84b0ea7851628ab17d8b0b3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12d60c8abecb3bc9b84b0ea7851628ab17d8b0b3.hip
deleted file mode 100644
index bc86062c2f2e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_12d60c8abecb3bc9b84b0ea7851628ab17d8b0b3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_131691f01cc7f29affb88152dd48c7a484315dcd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_131691f01cc7f29affb88152dd48c7a484315dcd.hip
deleted file mode 100644
index 7a562779069fb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_131691f01cc7f29affb88152dd48c7a484315dcd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_131c1fdc4206bb952b2fea675f24e3b09f605eef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_131c1fdc4206bb952b2fea675f24e3b09f605eef.hip
deleted file mode 100644
index b42587134ef0a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_131c1fdc4206bb952b2fea675f24e3b09f605eef.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_133c51948cf8584900807998da14d788039f53b9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_133c51948cf8584900807998da14d788039f53b9.hip
deleted file mode 100644
index 63e35572c92bb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_133c51948cf8584900807998da14d788039f53b9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_135ea67de101135ed5fe04f5cab1ec1d7b3714bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_135ea67de101135ed5fe04f5cab1ec1d7b3714bb.hip
deleted file mode 100644
index c071ea137ca23..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_135ea67de101135ed5fe04f5cab1ec1d7b3714bb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_137fa6780d9e6bde10aec10a875c039fdbbc652e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_137fa6780d9e6bde10aec10a875c039fdbbc652e.hip
deleted file mode 100644
index 679eb8d608964..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_137fa6780d9e6bde10aec10a875c039fdbbc652e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1386cd75411e61a8dbbaf2b916e62f4f5f99104f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1386cd75411e61a8dbbaf2b916e62f4f5f99104f.hip
deleted file mode 100644
index e7b2fc5cf1fe9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1386cd75411e61a8dbbaf2b916e62f4f5f99104f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_13d5f2ec83b3331654e37ea0b44d88cd98abaa37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_13d5f2ec83b3331654e37ea0b44d88cd98abaa37.hip
deleted file mode 100644
index 86d734a777187..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_13d5f2ec83b3331654e37ea0b44d88cd98abaa37.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_13f747525ad31e76c88774fb2208e470da9c2310.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_13f747525ad31e76c88774fb2208e470da9c2310.hip
deleted file mode 100644
index 5b29c2e23a96c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_13f747525ad31e76c88774fb2208e470da9c2310.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14221590b90c48d3cf259fb4e834ccfaf7f3209b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14221590b90c48d3cf259fb4e834ccfaf7f3209b.hip
deleted file mode 100644
index e554c56f71ff8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14221590b90c48d3cf259fb4e834ccfaf7f3209b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_144f19363ef26efd36f0436cfa9f84f181a8824c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_144f19363ef26efd36f0436cfa9f84f181a8824c.hip
deleted file mode 100644
index 238d01d41536d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_144f19363ef26efd36f0436cfa9f84f181a8824c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_146eb8c40e3146e06936f3141b2c4d92a578ddec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_146eb8c40e3146e06936f3141b2c4d92a578ddec.hip
deleted file mode 100644
index 42f40c02e6a86..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_146eb8c40e3146e06936f3141b2c4d92a578ddec.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14baaaf1e90a075ab802c6e7d97c4b1605c8bd72.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14baaaf1e90a075ab802c6e7d97c4b1605c8bd72.hip
deleted file mode 100644
index 2a2e0ba23003f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14baaaf1e90a075ab802c6e7d97c4b1605c8bd72.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14c4ebd1792c781d219bd21b691b575f64635730.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14c4ebd1792c781d219bd21b691b575f64635730.hip
deleted file mode 100644
index 68b15fe07b80e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14c4ebd1792c781d219bd21b691b575f64635730.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14d11aad7b666f500f68b264a2fcca6dfc5f1a05.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14d11aad7b666f500f68b264a2fcca6dfc5f1a05.hip
deleted file mode 100644
index d6827fe611b57..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14d11aad7b666f500f68b264a2fcca6dfc5f1a05.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14d4630876785655bd4950566e81ae0b645c0d3c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14d4630876785655bd4950566e81ae0b645c0d3c.hip
deleted file mode 100644
index 968af70552b1a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14d4630876785655bd4950566e81ae0b645c0d3c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14f77aeeafe4b28f314fde5ebccfd2a554872781.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14f77aeeafe4b28f314fde5ebccfd2a554872781.hip
deleted file mode 100644
index 15a81f690ef9e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14f77aeeafe4b28f314fde5ebccfd2a554872781.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14fea611f3c253aebf726af3e5fdb7e63e18e13a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14fea611f3c253aebf726af3e5fdb7e63e18e13a.hip
deleted file mode 100644
index e9281c2470bc6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_14fea611f3c253aebf726af3e5fdb7e63e18e13a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_151a4425b411596c46c7032f6b83d3152a0e0cd4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_151a4425b411596c46c7032f6b83d3152a0e0cd4.hip
deleted file mode 100644
index a9090f2c4a6f6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_151a4425b411596c46c7032f6b83d3152a0e0cd4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_153e897098539c3466da9d7a37234daf16476277.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_153e897098539c3466da9d7a37234daf16476277.hip
deleted file mode 100644
index b034d2a816dd6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_153e897098539c3466da9d7a37234daf16476277.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1552dc38d26f6badb7a9bcb5ce9124d54cc45ed3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1552dc38d26f6badb7a9bcb5ce9124d54cc45ed3.hip
deleted file mode 100644
index 89ecc02f153e2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1552dc38d26f6badb7a9bcb5ce9124d54cc45ed3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_155bafb551768855c8c01faa63e44764ebe6c110.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_155bafb551768855c8c01faa63e44764ebe6c110.hip
deleted file mode 100644
index ae9c12940f1a5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_155bafb551768855c8c01faa63e44764ebe6c110.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_155c3549d067464d186a99b8205317cc000d4898.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_155c3549d067464d186a99b8205317cc000d4898.hip
deleted file mode 100644
index 023d6d7a8fabf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_155c3549d067464d186a99b8205317cc000d4898.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1573e3d855d28c54af612ab950b081302891d56d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1573e3d855d28c54af612ab950b081302891d56d.hip
deleted file mode 100644
index 73c97ded48fe8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1573e3d855d28c54af612ab950b081302891d56d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_157768cd725813f8111d265cfdfea7f42034e5e9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_157768cd725813f8111d265cfdfea7f42034e5e9.hip
deleted file mode 100644
index b21f5b7f8d1a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_157768cd725813f8111d265cfdfea7f42034e5e9.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_157b89d8d625b8244b5cceaa4d3e5fc5a09c8989.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_157b89d8d625b8244b5cceaa4d3e5fc5a09c8989.hip
deleted file mode 100644
index d1f406f622a8f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_157b89d8d625b8244b5cceaa4d3e5fc5a09c8989.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_158d5ce564c3ae1eefb54e3d41dde2604560ef4a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_158d5ce564c3ae1eefb54e3d41dde2604560ef4a.hip
deleted file mode 100644
index e10ea61bbc208..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_158d5ce564c3ae1eefb54e3d41dde2604560ef4a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_159ee1f1b44d1a8fbaead65d8449413bb616d15e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_159ee1f1b44d1a8fbaead65d8449413bb616d15e.hip
deleted file mode 100644
index d0948fa907a19..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_159ee1f1b44d1a8fbaead65d8449413bb616d15e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15b255dde1a9d915e582ee2a83de7d83190c6a24.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15b255dde1a9d915e582ee2a83de7d83190c6a24.hip
deleted file mode 100644
index fc63e594997ea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15b255dde1a9d915e582ee2a83de7d83190c6a24.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15cf7068183421b141ed5d6e7fe902d06b6492a1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15cf7068183421b141ed5d6e7fe902d06b6492a1.hip
deleted file mode 100644
index 430795d7cac19..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15cf7068183421b141ed5d6e7fe902d06b6492a1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15dc02ea7e0908cf0bd48034f5a49debfaa36219.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15dc02ea7e0908cf0bd48034f5a49debfaa36219.hip
deleted file mode 100644
index 2c15174057a27..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15dc02ea7e0908cf0bd48034f5a49debfaa36219.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15e8e1ab8c63db96843054bb7a98d708ae6a9c44.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15e8e1ab8c63db96843054bb7a98d708ae6a9c44.hip
deleted file mode 100644
index 3be74e604c106..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15e8e1ab8c63db96843054bb7a98d708ae6a9c44.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15fe3e8f4add16a088fe44458353fa7c0c4f9658.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15fe3e8f4add16a088fe44458353fa7c0c4f9658.hip
deleted file mode 100644
index 29e5a61870714..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_15fe3e8f4add16a088fe44458353fa7c0c4f9658.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_16047b5544acef40e39932672cac6f562e200948.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_16047b5544acef40e39932672cac6f562e200948.hip
deleted file mode 100644
index 3b7a1342e0e90..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_16047b5544acef40e39932672cac6f562e200948.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1621507cf219fe608715d4e5bb6e5764022e2d61.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1621507cf219fe608715d4e5bb6e5764022e2d61.hip
deleted file mode 100644
index fa589b3c519e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1621507cf219fe608715d4e5bb6e5764022e2d61.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_162b0dfbe3f615b1d164290799b2457437a0044b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_162b0dfbe3f615b1d164290799b2457437a0044b.hip
deleted file mode 100644
index cc4ef8e9d2718..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_162b0dfbe3f615b1d164290799b2457437a0044b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_164a947a6c2ba83a5b1cb7074aee0bdac6c9c64e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_164a947a6c2ba83a5b1cb7074aee0bdac6c9c64e.hip
deleted file mode 100644
index f2a1616785cef..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_164a947a6c2ba83a5b1cb7074aee0bdac6c9c64e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_165dfb45658df8f1ae8dc0738ac9614740f2576c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_165dfb45658df8f1ae8dc0738ac9614740f2576c.hip
deleted file mode 100644
index 8cc1759f9df03..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_165dfb45658df8f1ae8dc0738ac9614740f2576c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_167f5328b035ed59a6f05dfee31edd704c4b07ee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_167f5328b035ed59a6f05dfee31edd704c4b07ee.hip
deleted file mode 100644
index b68e24323a1f3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_167f5328b035ed59a6f05dfee31edd704c4b07ee.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1687ddf65ce4ed2997583e20fee9f201e86633b3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1687ddf65ce4ed2997583e20fee9f201e86633b3.hip
deleted file mode 100644
index 573dea9236ec1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1687ddf65ce4ed2997583e20fee9f201e86633b3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_16f94f5c65c37624f5458c165daf83517d9e3c81.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_16f94f5c65c37624f5458c165daf83517d9e3c81.hip
deleted file mode 100644
index 715c37a9bf390..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_16f94f5c65c37624f5458c165daf83517d9e3c81.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_173c44dd85077e6b12dd06fdcf6b11ba349e1866.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_173c44dd85077e6b12dd06fdcf6b11ba349e1866.hip
deleted file mode 100644
index 5721349d162a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_173c44dd85077e6b12dd06fdcf6b11ba349e1866.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_17b9b96edda151072215502cc2b606bf1f6f0b03.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_17b9b96edda151072215502cc2b606bf1f6f0b03.hip
deleted file mode 100644
index 2fe640db9a109..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_17b9b96edda151072215502cc2b606bf1f6f0b03.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1847fef2c06ea581b0ab31af1cb0556c572696ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1847fef2c06ea581b0ab31af1cb0556c572696ad.hip
deleted file mode 100644
index 3ef88ce2cc065..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1847fef2c06ea581b0ab31af1cb0556c572696ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_187963e1969301abfa61d06afc97faea2bb4efb1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_187963e1969301abfa61d06afc97faea2bb4efb1.hip
deleted file mode 100644
index 2bc002ad61c14..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_187963e1969301abfa61d06afc97faea2bb4efb1.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1886d4bf54b3a4a9e093360998b2059b3c03d072.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1886d4bf54b3a4a9e093360998b2059b3c03d072.hip
deleted file mode 100644
index f46020c36692e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1886d4bf54b3a4a9e093360998b2059b3c03d072.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_188a70d526394e254274df95de0727850820326c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_188a70d526394e254274df95de0727850820326c.hip
deleted file mode 100644
index a35b92ebd3aad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_188a70d526394e254274df95de0727850820326c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1899e28aff2fb168cdc3af7132dd7fd09c2e1ced.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1899e28aff2fb168cdc3af7132dd7fd09c2e1ced.hip
deleted file mode 100644
index e4fe0ad72f1fe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1899e28aff2fb168cdc3af7132dd7fd09c2e1ced.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18a4d71b31c451a50df7996e3db864bc3c3882ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18a4d71b31c451a50df7996e3db864bc3c3882ed.hip
deleted file mode 100644
index a74f83d2c6cbc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18a4d71b31c451a50df7996e3db864bc3c3882ed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18b92b4e249195ac3e0c74d246585a4c9e0992fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18b92b4e249195ac3e0c74d246585a4c9e0992fd.hip
deleted file mode 100644
index f86e45c6e75f5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18b92b4e249195ac3e0c74d246585a4c9e0992fd.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18ed7195a9443c84956c3f32839cb3ab9056bdfc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18ed7195a9443c84956c3f32839cb3ab9056bdfc.hip
deleted file mode 100644
index 482ef07d6a85c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_18ed7195a9443c84956c3f32839cb3ab9056bdfc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1914250fce818584291c69a5f058a58cfbd83df9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1914250fce818584291c69a5f058a58cfbd83df9.hip
deleted file mode 100644
index 6d626566787a5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1914250fce818584291c69a5f058a58cfbd83df9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_193699a5daa14ca2def07489e0b563149bc403f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_193699a5daa14ca2def07489e0b563149bc403f8.hip
deleted file mode 100644
index 04fcedfa0e6c6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_193699a5daa14ca2def07489e0b563149bc403f8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19af6a7f9e5020e8d0f0ca0f6258001f6ce592c1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19af6a7f9e5020e8d0f0ca0f6258001f6ce592c1.hip
deleted file mode 100644
index 14a181fa2ca5d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19af6a7f9e5020e8d0f0ca0f6258001f6ce592c1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19cd9f7b08cec83736605af63d9fcaf463a1aea4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19cd9f7b08cec83736605af63d9fcaf463a1aea4.hip
deleted file mode 100644
index 35f450cf73e7f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19cd9f7b08cec83736605af63d9fcaf463a1aea4.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19df4e13108e043361e9528b71df56f04f696a0c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19df4e13108e043361e9528b71df56f04f696a0c.hip
deleted file mode 100644
index bd2b484ab1b0f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_19df4e13108e043361e9528b71df56f04f696a0c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a11dd5ebb989503a1c182684e7f247e2f8cd9c2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a11dd5ebb989503a1c182684e7f247e2f8cd9c2.hip
deleted file mode 100644
index 9298ada0dcaff..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a11dd5ebb989503a1c182684e7f247e2f8cd9c2.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a236be9da05a07d11cd28034d90cdf89941a172.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a236be9da05a07d11cd28034d90cdf89941a172.hip
deleted file mode 100644
index 025ff51c4bb54..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a236be9da05a07d11cd28034d90cdf89941a172.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a5e18f6333ed2cce509f07cb8bd5868951d66a0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a5e18f6333ed2cce509f07cb8bd5868951d66a0.hip
deleted file mode 100644
index 01fe775956b9b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a5e18f6333ed2cce509f07cb8bd5868951d66a0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a6785392af35e27d6697b584cb6f17a766d3fee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a6785392af35e27d6697b584cb6f17a766d3fee.hip
deleted file mode 100644
index b8076aeff3475..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a6785392af35e27d6697b584cb6f17a766d3fee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a6bc2762b95d550485aa720edaf71138d94cd07.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a6bc2762b95d550485aa720edaf71138d94cd07.hip
deleted file mode 100644
index 95adf9fe618e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a6bc2762b95d550485aa720edaf71138d94cd07.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a8da3e6ab050262b659c801ccf9a14787d7f176.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a8da3e6ab050262b659c801ccf9a14787d7f176.hip
deleted file mode 100644
index 4a426d3b44bd2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a8da3e6ab050262b659c801ccf9a14787d7f176.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a96f0ac76f117e66eba97cb990c2350561ec2ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a96f0ac76f117e66eba97cb990c2350561ec2ab.hip
deleted file mode 100644
index 47831996f5d0c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a96f0ac76f117e66eba97cb990c2350561ec2ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a98bcbe900f8c141136d18c114b02fffbe8bca1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a98bcbe900f8c141136d18c114b02fffbe8bca1.hip
deleted file mode 100644
index c7f2e83abcb75..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a98bcbe900f8c141136d18c114b02fffbe8bca1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a99b2625adffa8215276bb88fc65bae944b846b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a99b2625adffa8215276bb88fc65bae944b846b.hip
deleted file mode 100644
index 181a3ee0a95e3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1a99b2625adffa8215276bb88fc65bae944b846b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1acf2f892742b1d236d2b31a8185c6869126adad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1acf2f892742b1d236d2b31a8185c6869126adad.hip
deleted file mode 100644
index cf9f36d0a048a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1acf2f892742b1d236d2b31a8185c6869126adad.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1b3e7c8969027d3316875f33dc50fe022e05ce37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1b3e7c8969027d3316875f33dc50fe022e05ce37.hip
deleted file mode 100644
index 6d7d63c8f1661..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1b3e7c8969027d3316875f33dc50fe022e05ce37.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1be43f8b629e7039f57b95866d5777273377470d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1be43f8b629e7039f57b95866d5777273377470d.hip
deleted file mode 100644
index 47801621a594f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1be43f8b629e7039f57b95866d5777273377470d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1be746990a2032f0363ad9f9112cc994983f4706.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1be746990a2032f0363ad9f9112cc994983f4706.hip
deleted file mode 100644
index 248b8f51f98d9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1be746990a2032f0363ad9f9112cc994983f4706.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1bf767e7104cfc8322f26df35907fbf04b8948f3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1bf767e7104cfc8322f26df35907fbf04b8948f3.hip
deleted file mode 100644
index 0d1b800a323c5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1bf767e7104cfc8322f26df35907fbf04b8948f3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c1b0f85e085dd0769c566fb16aafe5ab5952714.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c1b0f85e085dd0769c566fb16aafe5ab5952714.hip
deleted file mode 100644
index 243f6df90cf84..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c1b0f85e085dd0769c566fb16aafe5ab5952714.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c2a2d78176e3f0a78e3ad78217e75a4430c0de5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c2a2d78176e3f0a78e3ad78217e75a4430c0de5.hip
deleted file mode 100644
index c47609b9a2b3b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c2a2d78176e3f0a78e3ad78217e75a4430c0de5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c65ba6dba01da9caa84ba89453b61d81376763f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c65ba6dba01da9caa84ba89453b61d81376763f.hip
deleted file mode 100644
index 95217f517ba50..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1c65ba6dba01da9caa84ba89453b61d81376763f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1ca3f45d0be2d1119cccd0af042a3e8adeda2ed7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1ca3f45d0be2d1119cccd0af042a3e8adeda2ed7.hip
deleted file mode 100644
index 30608a330e5a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1ca3f45d0be2d1119cccd0af042a3e8adeda2ed7.hip
+++ /dev/null
@@ -1,1965 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-float fmha_fwd(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){
-    float r = -1;
-    if(t.data_type.compare("fp16") == 0){
-        if (t.hdim_q <= 32 && t.hdim_v <= 32) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::fp16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-        else if (t.hdim_q <= 64 && t.hdim_v <= 64) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::fp16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-        else if (t.hdim_q <= 128 && t.hdim_v <= 128) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::fp16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-        else if (t.hdim_q <= 256 && t.hdim_v <= 256) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::fp16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-
-    }
-    else if(t.data_type.compare("bf16") == 0){
-        if (t.hdim_q <= 32 && t.hdim_v <= 32) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, false, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<32, ck_tile::bf16_t, true, 128, 64, 16, 32, 32, 32, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-        else if (t.hdim_q <= 64 && t.hdim_v <= 64) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 64 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 64 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, false, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<64, ck_tile::bf16_t, true, 128, 64, 32, 64, 32, 64, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-        else if (t.hdim_q <= 128 && t.hdim_v <= 128) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k != 0 && a.seqlen_k % 128 == 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true) && (a.seqlen_k == 0 || a.seqlen_k % 128 != 0) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, false, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (a.hdim_q % 8 == 0) && (a.hdim_v % 8 == 0)) {
-                using trait_ = fmha_fwd_traits_<128, ck_tile::bf16_t, true, 128, 128, 32, 128, 32, 128, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-        else if (t.hdim_q <= 256 && t.hdim_v <= 256) {
-            if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (a.seqlen_q % 128 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == false) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true /*a.seqlen_q % 128 != 0*/) && (true /*a.seqlen_k % 128 != 0*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, false, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == true)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == true) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-            else if((t.is_group_mode == true) && (t.is_v_rowmajor == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_lse == false)  && (t.has_dropout == false) && (t.do_fp8_static_quant == false) &&
-                        (true/*group mode spad always true*/) && (true/*group mode skpad always true*/) && (true /*a.hdim_q % 256 != 0*/) && (true /*a.hdim_v % 256 != 0*/)) {
-                using trait_ = fmha_fwd_traits_<256, ck_tile::bf16_t, true, 128, 128, 32, 256, 32, 256, true, ck_tile::BlockFmhaPipelineEnum::QRKSVS, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-                return fmha_fwd_<trait_>(s, a);
-            }
-
-        }
-
-    }
-
-    return r;
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1cbf88db44aa5f884438288a325270d29c7a04b6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1cbf88db44aa5f884438288a325270d29c7a04b6.hip
deleted file mode 100644
index 73391c7122655..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1cbf88db44aa5f884438288a325270d29c7a04b6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1cc459e57bfed5ec7f40ea4a4dd9f72f3ad7a709.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1cc459e57bfed5ec7f40ea4a4dd9f72f3ad7a709.hip
deleted file mode 100644
index f482fd18d2a13..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1cc459e57bfed5ec7f40ea4a4dd9f72f3ad7a709.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d02609fb803ea2697e2c2cef35e6f923d2578cf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d02609fb803ea2697e2c2cef35e6f923d2578cf.hip
deleted file mode 100644
index 4e5b46371b7c6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d02609fb803ea2697e2c2cef35e6f923d2578cf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d0b822743e0205f60521d38d7c64f589fdf0f58.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d0b822743e0205f60521d38d7c64f589fdf0f58.hip
deleted file mode 100644
index 836aa4048ee2a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d0b822743e0205f60521d38d7c64f589fdf0f58.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d21263e16dafe79b9fe2f998847296e575c14e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d21263e16dafe79b9fe2f998847296e575c14e7.hip
deleted file mode 100644
index fb779faa4a2db..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d21263e16dafe79b9fe2f998847296e575c14e7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d3ef3d5ded0dfe2a0bafb52ea8f841658db35fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d3ef3d5ded0dfe2a0bafb52ea8f841658db35fd.hip
deleted file mode 100644
index 94eff94b23c80..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d3ef3d5ded0dfe2a0bafb52ea8f841658db35fd.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d498e418ebbf33bed58b4074d1edf3d9bdd07c5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d498e418ebbf33bed58b4074d1edf3d9bdd07c5.hip
deleted file mode 100644
index 5a57720770aaa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1d498e418ebbf33bed58b4074d1edf3d9bdd07c5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1da23de9604b5d98fe02529075bad995954c12ca.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1da23de9604b5d98fe02529075bad995954c12ca.hip
deleted file mode 100644
index e9f2586443c00..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1da23de9604b5d98fe02529075bad995954c12ca.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1db03461737f1e359f389a8d297476f9b60faabd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1db03461737f1e359f389a8d297476f9b60faabd.hip
deleted file mode 100644
index a6baa018aff3e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1db03461737f1e359f389a8d297476f9b60faabd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1dc6e599144a093203fd7f92ac6d3c2cd7180d49.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1dc6e599144a093203fd7f92ac6d3c2cd7180d49.hip
deleted file mode 100644
index cf3d8876d6219..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1dc6e599144a093203fd7f92ac6d3c2cd7180d49.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1de2f97d49f015b9af0b186801e939c6f357a0c4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1de2f97d49f015b9af0b186801e939c6f357a0c4.hip
deleted file mode 100644
index 6925977e633fb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1de2f97d49f015b9af0b186801e939c6f357a0c4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1df893ee660d37fba7eaca452ae65b3e45a73087.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1df893ee660d37fba7eaca452ae65b3e45a73087.hip
deleted file mode 100644
index ceefbd5e8fcb8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1df893ee660d37fba7eaca452ae65b3e45a73087.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e22f2d99804198c61251b4629a3f18ed3dcd42e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e22f2d99804198c61251b4629a3f18ed3dcd42e.hip
deleted file mode 100644
index c6b5424511052..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e22f2d99804198c61251b4629a3f18ed3dcd42e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e33ce1fa113b221e5303b4093c2c4e748ce8298.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e33ce1fa113b221e5303b4093c2c4e748ce8298.hip
deleted file mode 100644
index 4a2cdbc008b87..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e33ce1fa113b221e5303b4093c2c4e748ce8298.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e42736d4f677a59a172bd6f162616a437696351.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e42736d4f677a59a172bd6f162616a437696351.hip
deleted file mode 100644
index 5c478d32afe59..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e42736d4f677a59a172bd6f162616a437696351.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e7d7888480b83c78833214b32e10f37a6e20301.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e7d7888480b83c78833214b32e10f37a6e20301.hip
deleted file mode 100644
index 60f771fac945c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e7d7888480b83c78833214b32e10f37a6e20301.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e9130607a2d24cb0662a47e9cf12c6602143838.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e9130607a2d24cb0662a47e9cf12c6602143838.hip
deleted file mode 100644
index d66b60ea57917..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e9130607a2d24cb0662a47e9cf12c6602143838.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e943fcc2e64c618fc1415b3f1a0db4d70aa8494.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e943fcc2e64c618fc1415b3f1a0db4d70aa8494.hip
deleted file mode 100644
index 9164d233343e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1e943fcc2e64c618fc1415b3f1a0db4d70aa8494.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1edaf9d4270d2ac61c299320e06ba73f44730364.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1edaf9d4270d2ac61c299320e06ba73f44730364.hip
deleted file mode 100644
index 89688d0ff5e8b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1edaf9d4270d2ac61c299320e06ba73f44730364.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f0cad6ad5b172e51c569e84cd54a19b4eb0ed05.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f0cad6ad5b172e51c569e84cd54a19b4eb0ed05.hip
deleted file mode 100644
index df15d8e39bced..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f0cad6ad5b172e51c569e84cd54a19b4eb0ed05.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f13a6d0f8c798c0c4ba4ad202d081899fe081ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f13a6d0f8c798c0c4ba4ad202d081899fe081ab.hip
deleted file mode 100644
index c0e8df1f9478e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f13a6d0f8c798c0c4ba4ad202d081899fe081ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f6bc5faf18be193212217788d476ce6fd384bfb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f6bc5faf18be193212217788d476ce6fd384bfb.hip
deleted file mode 100644
index 93513afa554eb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f6bc5faf18be193212217788d476ce6fd384bfb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f7faa0b33a9aada86f032174afd40d18efa7715.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f7faa0b33a9aada86f032174afd40d18efa7715.hip
deleted file mode 100644
index 6b6a858e06dda..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f7faa0b33a9aada86f032174afd40d18efa7715.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f81f8cce0d77dec9f977b9eeb0778b70a13fa75.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f81f8cce0d77dec9f977b9eeb0778b70a13fa75.hip
deleted file mode 100644
index 9e1dcb2e6f890..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1f81f8cce0d77dec9f977b9eeb0778b70a13fa75.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fcdcb750f382fc7828a9886585f50efbe5be735.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fcdcb750f382fc7828a9886585f50efbe5be735.hip
deleted file mode 100644
index aea7ec22eca11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fcdcb750f382fc7828a9886585f50efbe5be735.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fd9fa7c2e13d0bad5fddb2b5a316bbc09d397ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fd9fa7c2e13d0bad5fddb2b5a316bbc09d397ea.hip
deleted file mode 100644
index 334483fc2439a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fd9fa7c2e13d0bad5fddb2b5a316bbc09d397ea.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fda1c96568eab89a8f6498f8bb23c1223cdc7b0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fda1c96568eab89a8f6498f8bb23c1223cdc7b0.hip
deleted file mode 100644
index efa3ae279800a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_1fda1c96568eab89a8f6498f8bb23c1223cdc7b0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2005aca3520b171bb82d10ad70fef44f28c19776.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2005aca3520b171bb82d10ad70fef44f28c19776.hip
deleted file mode 100644
index 8041b74da8525..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2005aca3520b171bb82d10ad70fef44f28c19776.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_204a573ce6b7d2f90aede543939315561cc43177.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_204a573ce6b7d2f90aede543939315561cc43177.hip
deleted file mode 100644
index 09c17b863cd4f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_204a573ce6b7d2f90aede543939315561cc43177.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20588bcac681a5d69f252d7523a3681a0c6b6181.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20588bcac681a5d69f252d7523a3681a0c6b6181.hip
deleted file mode 100644
index 865812b92a157..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20588bcac681a5d69f252d7523a3681a0c6b6181.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2081430c92864c29bb9f409e7c27caee1de00749.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2081430c92864c29bb9f409e7c27caee1de00749.hip
deleted file mode 100644
index f963a04e71b5d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2081430c92864c29bb9f409e7c27caee1de00749.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20d5c3c86398f6ce55abc90db3e362dbf9f457f2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20d5c3c86398f6ce55abc90db3e362dbf9f457f2.hip
deleted file mode 100644
index c332023863c82..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20d5c3c86398f6ce55abc90db3e362dbf9f457f2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20f7ea0aabd069362ba4bbd66623cea5b6e1a6bd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20f7ea0aabd069362ba4bbd66623cea5b6e1a6bd.hip
deleted file mode 100644
index b8235acd7300e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_20f7ea0aabd069362ba4bbd66623cea5b6e1a6bd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_210ef512b7862837f54acbc3b21e135a192647a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_210ef512b7862837f54acbc3b21e135a192647a3.hip
deleted file mode 100644
index 37fcd87b8fda5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_210ef512b7862837f54acbc3b21e135a192647a3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2122c973581930ab7a4ebc90b3bf1cdaa229a87f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2122c973581930ab7a4ebc90b3bf1cdaa229a87f.hip
deleted file mode 100644
index d88ddc61be7cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2122c973581930ab7a4ebc90b3bf1cdaa229a87f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21411df58165946bf02942b597d94de7dd856987.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21411df58165946bf02942b597d94de7dd856987.hip
deleted file mode 100644
index 671921ec00fe7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21411df58165946bf02942b597d94de7dd856987.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_216806a4598c885e517e664fc8280c59ec3cbf11.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_216806a4598c885e517e664fc8280c59ec3cbf11.hip
deleted file mode 100644
index 91b2de03324eb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_216806a4598c885e517e664fc8280c59ec3cbf11.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2173b7c710d418f44dc2b41bec5905024334eae5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2173b7c710d418f44dc2b41bec5905024334eae5.hip
deleted file mode 100644
index 5fe4ecad692c8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2173b7c710d418f44dc2b41bec5905024334eae5.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2177d95cdf45f6fec95d1812f2ef183a75259e38.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2177d95cdf45f6fec95d1812f2ef183a75259e38.hip
deleted file mode 100644
index 8f2a0a6e6eaa9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2177d95cdf45f6fec95d1812f2ef183a75259e38.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21828c7d3f5574690f12f841c27f025206e6165b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21828c7d3f5574690f12f841c27f025206e6165b.hip
deleted file mode 100644
index bb6073dc6fc99..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21828c7d3f5574690f12f841c27f025206e6165b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2184fba2eec5899bb40d49d4508196e6be1ec1b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2184fba2eec5899bb40d49d4508196e6be1ec1b1.hip
deleted file mode 100644
index 58fc638b835cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2184fba2eec5899bb40d49d4508196e6be1ec1b1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21e235e31d6955393ac8e825bd69ead70687b7c8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21e235e31d6955393ac8e825bd69ead70687b7c8.hip
deleted file mode 100644
index be18058190aab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21e235e31d6955393ac8e825bd69ead70687b7c8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21f860d42fdc2cc6bd743d53ba546e332c22fedf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21f860d42fdc2cc6bd743d53ba546e332c22fedf.hip
deleted file mode 100644
index 27699ac75a739..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_21f860d42fdc2cc6bd743d53ba546e332c22fedf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22105635385fbfb5d2f330df83ba6747bcb27f6d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22105635385fbfb5d2f330df83ba6747bcb27f6d.hip
deleted file mode 100644
index b894a81ea9f26..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22105635385fbfb5d2f330df83ba6747bcb27f6d.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_224f9af5e5ca519b21b71a54acb49f50b4999c47.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_224f9af5e5ca519b21b71a54acb49f50b4999c47.hip
deleted file mode 100644
index 5bd291bb9f846..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_224f9af5e5ca519b21b71a54acb49f50b4999c47.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22511de2592b6e350737e44865e1fed6496e3f32.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22511de2592b6e350737e44865e1fed6496e3f32.hip
deleted file mode 100644
index 83689a58fa0ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22511de2592b6e350737e44865e1fed6496e3f32.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22632f996eb63fbe4bc5748c5897b775087446a0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22632f996eb63fbe4bc5748c5897b775087446a0.hip
deleted file mode 100644
index 8c312dca66c0e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22632f996eb63fbe4bc5748c5897b775087446a0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_226662cf1c9900a4334d2cadcc5f5ac3ad355f05.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_226662cf1c9900a4334d2cadcc5f5ac3ad355f05.hip
deleted file mode 100644
index 9003c098b391e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_226662cf1c9900a4334d2cadcc5f5ac3ad355f05.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2273457ac3be01cc1595a015a5f598f8290c77e4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2273457ac3be01cc1595a015a5f598f8290c77e4.hip
deleted file mode 100644
index e11c6d0aca7c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2273457ac3be01cc1595a015a5f598f8290c77e4.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22a07ecf1a59f72ec6bef3e970d7f33cf54c5f44.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22a07ecf1a59f72ec6bef3e970d7f33cf54c5f44.hip
deleted file mode 100644
index 1d8b60736452e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22a07ecf1a59f72ec6bef3e970d7f33cf54c5f44.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22c142d869ef940ca876c93033ad53b576ed34f2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22c142d869ef940ca876c93033ad53b576ed34f2.hip
deleted file mode 100644
index ee6ff72badae9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_22c142d869ef940ca876c93033ad53b576ed34f2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23047ea90076e3b0a3eb0586d49b9ee74ca6d279.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23047ea90076e3b0a3eb0586d49b9ee74ca6d279.hip
deleted file mode 100644
index b7c815e3791b8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23047ea90076e3b0a3eb0586d49b9ee74ca6d279.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_230861e81e5acc523fa680534eed757b7b4a4e1d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_230861e81e5acc523fa680534eed757b7b4a4e1d.hip
deleted file mode 100644
index afd118c1bbdfa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_230861e81e5acc523fa680534eed757b7b4a4e1d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_232f61bf31dbb5de5d7039d5ff2338068a759b68.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_232f61bf31dbb5de5d7039d5ff2338068a759b68.hip
deleted file mode 100644
index b4eef334b9ae0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_232f61bf31dbb5de5d7039d5ff2338068a759b68.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_233132e712eba8972ba444c604f89e01c5b84cc0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_233132e712eba8972ba444c604f89e01c5b84cc0.hip
deleted file mode 100644
index 5ce878e29280b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_233132e712eba8972ba444c604f89e01c5b84cc0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_235bf652702c2976551778b9159e09188575c63c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_235bf652702c2976551778b9159e09188575c63c.hip
deleted file mode 100644
index 5c69cd3a6eeb9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_235bf652702c2976551778b9159e09188575c63c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_236b3eef02b904304348b9d35f715b639d63218f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_236b3eef02b904304348b9d35f715b639d63218f.hip
deleted file mode 100644
index 16af5ff2193e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_236b3eef02b904304348b9d35f715b639d63218f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_238e4c1ca112afec494fbe47a85b553302c43395.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_238e4c1ca112afec494fbe47a85b553302c43395.hip
deleted file mode 100644
index 136c6cfa2e4af..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_238e4c1ca112afec494fbe47a85b553302c43395.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23914c00690ac5c4f89cdbbaf00732ba66c5c0ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23914c00690ac5c4f89cdbbaf00732ba66c5c0ef.hip
deleted file mode 100644
index e3dd044de7d45..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23914c00690ac5c4f89cdbbaf00732ba66c5c0ef.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23c9b46da8774462de8c24e14b12df3ed596eb57.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23c9b46da8774462de8c24e14b12df3ed596eb57.hip
deleted file mode 100644
index 4eaff6e7b5507..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_23c9b46da8774462de8c24e14b12df3ed596eb57.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_242013527a0266ad479715ee3e6ae01c45de29d0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_242013527a0266ad479715ee3e6ae01c45de29d0.hip
deleted file mode 100644
index 5fbd1baa08607..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_242013527a0266ad479715ee3e6ae01c45de29d0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_24410fd9a4150c33186a2a365d06d8f6ea621c20.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_24410fd9a4150c33186a2a365d06d8f6ea621c20.hip
deleted file mode 100644
index 77b4c86338366..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_24410fd9a4150c33186a2a365d06d8f6ea621c20.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_245d90000b55ab8b6055b1934880fc6c4870b34b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_245d90000b55ab8b6055b1934880fc6c4870b34b.hip
deleted file mode 100644
index 28691cb2b5c1a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_245d90000b55ab8b6055b1934880fc6c4870b34b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_24643917fc970c043d1c80d8d4b17ec92deeb8a1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_24643917fc970c043d1c80d8d4b17ec92deeb8a1.hip
deleted file mode 100644
index ca3e3d45b7aed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_24643917fc970c043d1c80d8d4b17ec92deeb8a1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_249668a3212cd00edaae871758be30a5a1fea589.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_249668a3212cd00edaae871758be30a5a1fea589.hip
deleted file mode 100644
index c9634cbab81c8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_249668a3212cd00edaae871758be30a5a1fea589.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_249e6b93baae25dff97a0bc9145a8d328ed3f317.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_249e6b93baae25dff97a0bc9145a8d328ed3f317.hip
deleted file mode 100644
index 20d06011bc1b9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_249e6b93baae25dff97a0bc9145a8d328ed3f317.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2543da478310245e19e6c6a0d9ed7ad99540b3bc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2543da478310245e19e6c6a0d9ed7ad99540b3bc.hip
deleted file mode 100644
index 8a21b5c90defd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2543da478310245e19e6c6a0d9ed7ad99540b3bc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_256ef175029a43e64164176d4eb212baf9d27bb9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_256ef175029a43e64164176d4eb212baf9d27bb9.hip
deleted file mode 100644
index 7480ed6987e7f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_256ef175029a43e64164176d4eb212baf9d27bb9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_258d747083272ea657604ac84867ecea17bd65da.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_258d747083272ea657604ac84867ecea17bd65da.hip
deleted file mode 100644
index 5e57c6c6d3ee8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_258d747083272ea657604ac84867ecea17bd65da.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_25938733446b6c0dcd159719f08d04a9aa467967.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_25938733446b6c0dcd159719f08d04a9aa467967.hip
deleted file mode 100644
index 783157a28e5a3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_25938733446b6c0dcd159719f08d04a9aa467967.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_25b3225da1e1842f83592971a1f62a0fe30aa9d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_25b3225da1e1842f83592971a1f62a0fe30aa9d3.hip
deleted file mode 100644
index 855839f90bd33..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_25b3225da1e1842f83592971a1f62a0fe30aa9d3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2660282ad39ef034fecbdb74acedfb48620b7dfd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2660282ad39ef034fecbdb74acedfb48620b7dfd.hip
deleted file mode 100644
index b147553021e01..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2660282ad39ef034fecbdb74acedfb48620b7dfd.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26835ba70606c769e56d19dbfe74061361aa855e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26835ba70606c769e56d19dbfe74061361aa855e.hip
deleted file mode 100644
index a43a8c06c9079..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26835ba70606c769e56d19dbfe74061361aa855e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2695783ae8f0034692efd6563f789ef03fd0f4f3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2695783ae8f0034692efd6563f789ef03fd0f4f3.hip
deleted file mode 100644
index e83e4ad89fa81..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2695783ae8f0034692efd6563f789ef03fd0f4f3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26d77b228420a3ead919474ec9c6fb2800f86890.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26d77b228420a3ead919474ec9c6fb2800f86890.hip
deleted file mode 100644
index 43d564ac04a54..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26d77b228420a3ead919474ec9c6fb2800f86890.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26ea90eb5a527434c1740933a1d2dd863eccf14c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26ea90eb5a527434c1740933a1d2dd863eccf14c.hip
deleted file mode 100644
index d3c3953617b9e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26ea90eb5a527434c1740933a1d2dd863eccf14c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26f90358e522d7bb7c76c3a2c6010f0f38788bb6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26f90358e522d7bb7c76c3a2c6010f0f38788bb6.hip
deleted file mode 100644
index a2b6a74abf636..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_26f90358e522d7bb7c76c3a2c6010f0f38788bb6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2703018e71d57d3266fc35e2e18a78faa3dd52ce.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2703018e71d57d3266fc35e2e18a78faa3dd52ce.hip
deleted file mode 100644
index c14490fad2990..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2703018e71d57d3266fc35e2e18a78faa3dd52ce.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_278639d44a4a8372a627a7c31e9527c8faa26f97.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_278639d44a4a8372a627a7c31e9527c8faa26f97.hip
deleted file mode 100644
index 2a5ec4564ac37..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_278639d44a4a8372a627a7c31e9527c8faa26f97.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_27c2000d32c230a57a6712f27bc0fba02722f5fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_27c2000d32c230a57a6712f27bc0fba02722f5fd.hip
deleted file mode 100644
index db00c82b5dd0e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_27c2000d32c230a57a6712f27bc0fba02722f5fd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_280bfced8745fbd9266207463fb41476dc23afff.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_280bfced8745fbd9266207463fb41476dc23afff.hip
deleted file mode 100644
index 1f63eb6826a27..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_280bfced8745fbd9266207463fb41476dc23afff.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_281d897ad17d7f6db2741b396e6b85a9b8f35286.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_281d897ad17d7f6db2741b396e6b85a9b8f35286.hip
deleted file mode 100644
index a501ac769e37c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_281d897ad17d7f6db2741b396e6b85a9b8f35286.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_285e61dad8f63fb973cb2eb899c959e400622652.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_285e61dad8f63fb973cb2eb899c959e400622652.hip
deleted file mode 100644
index f7fa0a6b2db2e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_285e61dad8f63fb973cb2eb899c959e400622652.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_288458c5a0720ef152848713119ebce6d76db6d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_288458c5a0720ef152848713119ebce6d76db6d6.hip
deleted file mode 100644
index 5b20a73750739..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_288458c5a0720ef152848713119ebce6d76db6d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_289071756e7d0582eb61ce6483fa3c988d2e10b5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_289071756e7d0582eb61ce6483fa3c988d2e10b5.hip
deleted file mode 100644
index 143b3336918ec..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_289071756e7d0582eb61ce6483fa3c988d2e10b5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28e4d2c757e4b8c366a2c320360e21ff0ef671a8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28e4d2c757e4b8c366a2c320360e21ff0ef671a8.hip
deleted file mode 100644
index 7022269ae247c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28e4d2c757e4b8c366a2c320360e21ff0ef671a8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f1ef32c4384ec26f3dc5e3af6a74fc8cebae92.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f1ef32c4384ec26f3dc5e3af6a74fc8cebae92.hip
deleted file mode 100644
index 59fdeb64a10ae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f1ef32c4384ec26f3dc5e3af6a74fc8cebae92.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f2e2b108a53308a0cb6c123c8d318cbc2eadb4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f2e2b108a53308a0cb6c123c8d318cbc2eadb4.hip
deleted file mode 100644
index 32fd3722d39a4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f2e2b108a53308a0cb6c123c8d318cbc2eadb4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f7634d29bef11fd466b452a46b0612f38c949b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f7634d29bef11fd466b452a46b0612f38c949b.hip
deleted file mode 100644
index ab520e38d6458..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_28f7634d29bef11fd466b452a46b0612f38c949b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_290c484c2a366258941ee0051e139ea716a9de2f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_290c484c2a366258941ee0051e139ea716a9de2f.hip
deleted file mode 100644
index 135ce1fc6fb6e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_290c484c2a366258941ee0051e139ea716a9de2f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_291a8bdf9d63b112e7fe5fa7e8835a6789cb8ecf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_291a8bdf9d63b112e7fe5fa7e8835a6789cb8ecf.hip
deleted file mode 100644
index 3d559a09fb650..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_291a8bdf9d63b112e7fe5fa7e8835a6789cb8ecf.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_292454f2d82184ab0491ea0675750c6ec55d659c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_292454f2d82184ab0491ea0675750c6ec55d659c.hip
deleted file mode 100644
index b0307da1caad0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_292454f2d82184ab0491ea0675750c6ec55d659c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_292b4f995d622826af5d1f2bffa7ba68467c841a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_292b4f995d622826af5d1f2bffa7ba68467c841a.hip
deleted file mode 100644
index 306043b0f0805..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_292b4f995d622826af5d1f2bffa7ba68467c841a.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_295a523f815eb822d66162d4feb75fe0bc50b648.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_295a523f815eb822d66162d4feb75fe0bc50b648.hip
deleted file mode 100644
index 1bb39b2a4f68c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_295a523f815eb822d66162d4feb75fe0bc50b648.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_296c5836ba118969c4ba89ed62a98dffe3105738.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_296c5836ba118969c4ba89ed62a98dffe3105738.hip
deleted file mode 100644
index 54214111437f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_296c5836ba118969c4ba89ed62a98dffe3105738.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2995d39cd62f20622a31f11a292ed175abb5fdf9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2995d39cd62f20622a31f11a292ed175abb5fdf9.hip
deleted file mode 100644
index d8a77346024f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2995d39cd62f20622a31f11a292ed175abb5fdf9.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29bffc159b0bb826ba489ae763dae141bfe8e802.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29bffc159b0bb826ba489ae763dae141bfe8e802.hip
deleted file mode 100644
index 26f9d5de83e62..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29bffc159b0bb826ba489ae763dae141bfe8e802.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29c9e5384809b21f39e78bb2e43af345a9a21d19.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29c9e5384809b21f39e78bb2e43af345a9a21d19.hip
deleted file mode 100644
index 87538d2c641a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29c9e5384809b21f39e78bb2e43af345a9a21d19.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29fe68ba10b3480dddc9866c51ca8b5efe962cc3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29fe68ba10b3480dddc9866c51ca8b5efe962cc3.hip
deleted file mode 100644
index 82ddd8a6332d1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_29fe68ba10b3480dddc9866c51ca8b5efe962cc3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a3a980a26682d879c3a3425f3ba5be3f5761adf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a3a980a26682d879c3a3425f3ba5be3f5761adf.hip
deleted file mode 100644
index 01b9f04a43c4f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a3a980a26682d879c3a3425f3ba5be3f5761adf.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a45129fc4995abcb8f880692f11c6186fc01641.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a45129fc4995abcb8f880692f11c6186fc01641.hip
deleted file mode 100644
index fc6d7c14ada43..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a45129fc4995abcb8f880692f11c6186fc01641.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a833fc01e88bd8e256ef64ae8251dd0ed10720b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a833fc01e88bd8e256ef64ae8251dd0ed10720b.hip
deleted file mode 100644
index efcc2d280d786..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a833fc01e88bd8e256ef64ae8251dd0ed10720b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a97c457144cb63a9c6c3d6be613b47bd0df9928.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a97c457144cb63a9c6c3d6be613b47bd0df9928.hip
deleted file mode 100644
index dafe4de3f38af..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2a97c457144cb63a9c6c3d6be613b47bd0df9928.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ad492377add5c8f6d0d2dbf9ee9e4338bbd9f1f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ad492377add5c8f6d0d2dbf9ee9e4338bbd9f1f.hip
deleted file mode 100644
index f2f1fccea3e3f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ad492377add5c8f6d0d2dbf9ee9e4338bbd9f1f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ae344010d49f7f9a6caab2cb84be7f87d2d96bf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ae344010d49f7f9a6caab2cb84be7f87d2d96bf.hip
deleted file mode 100644
index d37e3f4d1772d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ae344010d49f7f9a6caab2cb84be7f87d2d96bf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2af6c5be53732eb1939a2f93232af7dc011dec1a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2af6c5be53732eb1939a2f93232af7dc011dec1a.hip
deleted file mode 100644
index 433c8f0cb5cfe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2af6c5be53732eb1939a2f93232af7dc011dec1a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b0bcb241e5a1be1d35366461408d06e095a26ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b0bcb241e5a1be1d35366461408d06e095a26ef.hip
deleted file mode 100644
index 093825411226f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b0bcb241e5a1be1d35366461408d06e095a26ef.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b3326e055da32cc979892a2fbd0f7b003cb9f98.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b3326e055da32cc979892a2fbd0f7b003cb9f98.hip
deleted file mode 100644
index 67290678afc41..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b3326e055da32cc979892a2fbd0f7b003cb9f98.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b3af90387f1d227119c5dcd4b71362940bbce52.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b3af90387f1d227119c5dcd4b71362940bbce52.hip
deleted file mode 100644
index ff6efd764f80d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b3af90387f1d227119c5dcd4b71362940bbce52.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b4050988e5790a28dbe10b4c20e14f10f6cf85c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b4050988e5790a28dbe10b4c20e14f10f6cf85c.hip
deleted file mode 100644
index 96d0bd5d348cc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b4050988e5790a28dbe10b4c20e14f10f6cf85c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b49a9b0801a06dd89c7f7182d7590b515df1592.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b49a9b0801a06dd89c7f7182d7590b515df1592.hip
deleted file mode 100644
index ce052c1f899ee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b49a9b0801a06dd89c7f7182d7590b515df1592.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b50073f6dfeb7ea77d5dce288a1d2f08f8f6362.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b50073f6dfeb7ea77d5dce288a1d2f08f8f6362.hip
deleted file mode 100644
index 4381e1bb9c109..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b50073f6dfeb7ea77d5dce288a1d2f08f8f6362.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b5317b6cde327a842170ebff20c2b03d81379ff.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b5317b6cde327a842170ebff20c2b03d81379ff.hip
deleted file mode 100644
index d6118d7947565..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b5317b6cde327a842170ebff20c2b03d81379ff.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b8169ce4b4b9a17ac96fbb232e6a93f22071ab4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b8169ce4b4b9a17ac96fbb232e6a93f22071ab4.hip
deleted file mode 100644
index d2c2c39e21f5d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b8169ce4b4b9a17ac96fbb232e6a93f22071ab4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b823c3b99e7c8d1cdc39a5dbc7365a383bf9ccb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b823c3b99e7c8d1cdc39a5dbc7365a383bf9ccb.hip
deleted file mode 100644
index 8747df7c06ca1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2b823c3b99e7c8d1cdc39a5dbc7365a383bf9ccb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ba934408c75da5479cc41f96b98ea7d333635ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ba934408c75da5479cc41f96b98ea7d333635ea.hip
deleted file mode 100644
index fee242089c6c9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ba934408c75da5479cc41f96b98ea7d333635ea.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2bb6da1095bd8669c0e48b5cd808cf0dcefa2674.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2bb6da1095bd8669c0e48b5cd808cf0dcefa2674.hip
deleted file mode 100644
index fc42bd2c5183d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2bb6da1095bd8669c0e48b5cd808cf0dcefa2674.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c0bda0feaade2b554d648d72f219ac9c389bf09.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c0bda0feaade2b554d648d72f219ac9c389bf09.hip
deleted file mode 100644
index d565c77b50ef8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c0bda0feaade2b554d648d72f219ac9c389bf09.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c2e75e6f659a500dd3cf2cfd65118f111342119.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c2e75e6f659a500dd3cf2cfd65118f111342119.hip
deleted file mode 100644
index 878e0d921d3fd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c2e75e6f659a500dd3cf2cfd65118f111342119.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c77bd7e89ed832cc31b2995566a49bec6e4cb52.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c77bd7e89ed832cc31b2995566a49bec6e4cb52.hip
deleted file mode 100644
index dc1d525a87887..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c77bd7e89ed832cc31b2995566a49bec6e4cb52.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c7aede7762a524a7a424cc4dc46e43fdedf73a2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c7aede7762a524a7a424cc4dc46e43fdedf73a2.hip
deleted file mode 100644
index d2d7919946d24..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c7aede7762a524a7a424cc4dc46e43fdedf73a2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c808da5c2514806c2953bb77d5692e5d7c97aa3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c808da5c2514806c2953bb77d5692e5d7c97aa3.hip
deleted file mode 100644
index cb3deca9ea0be..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c808da5c2514806c2953bb77d5692e5d7c97aa3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c82e3c4e445e1e02f14435e4ca01a90850139a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c82e3c4e445e1e02f14435e4ca01a90850139a4.hip
deleted file mode 100644
index aea19e6c1a5a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c82e3c4e445e1e02f14435e4ca01a90850139a4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c9756060ac0e73dbcfc58a9222a78f0283cd029.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c9756060ac0e73dbcfc58a9222a78f0283cd029.hip
deleted file mode 100644
index 8ee7f8a3f76f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2c9756060ac0e73dbcfc58a9222a78f0283cd029.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2caba3ab83239e474412fcf89fe0fbef97e51bf1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2caba3ab83239e474412fcf89fe0fbef97e51bf1.hip
deleted file mode 100644
index 397ec87054787..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2caba3ab83239e474412fcf89fe0fbef97e51bf1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2cf351fc2c2da4a8e1760a3affc9a5947c6b3bda.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2cf351fc2c2da4a8e1760a3affc9a5947c6b3bda.hip
deleted file mode 100644
index adaeaee2d3e6b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2cf351fc2c2da4a8e1760a3affc9a5947c6b3bda.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d06f77a4054ca615d96636c0e2eba2a89850142.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d06f77a4054ca615d96636c0e2eba2a89850142.hip
deleted file mode 100644
index 8c8b346e59a1e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d06f77a4054ca615d96636c0e2eba2a89850142.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d1f2d1e57095f756ddd11e8e9d4f6f253e3ffa3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d1f2d1e57095f756ddd11e8e9d4f6f253e3ffa3.hip
deleted file mode 100644
index 900b96883a0aa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d1f2d1e57095f756ddd11e8e9d4f6f253e3ffa3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d23a26e0a59a8323dd97632e610d24624143fbe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d23a26e0a59a8323dd97632e610d24624143fbe.hip
deleted file mode 100644
index bd482b70420cc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d23a26e0a59a8323dd97632e610d24624143fbe.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d43460c011b8d5e01ea98c9b8ddce962de59a96.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d43460c011b8d5e01ea98c9b8ddce962de59a96.hip
deleted file mode 100644
index f2c62b6037c03..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d43460c011b8d5e01ea98c9b8ddce962de59a96.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d446754d7000673779d15d3e73039fd3c10a720.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d446754d7000673779d15d3e73039fd3c10a720.hip
deleted file mode 100644
index 63b3c94801472..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d446754d7000673779d15d3e73039fd3c10a720.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d7b637e0313cb423b22cd8844cc2997b3ff73e4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d7b637e0313cb423b22cd8844cc2997b3ff73e4.hip
deleted file mode 100644
index 610af33793e77..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d7b637e0313cb423b22cd8844cc2997b3ff73e4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d9a04b7f41dd6f0db017157a44790f35c626e2d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d9a04b7f41dd6f0db017157a44790f35c626e2d.hip
deleted file mode 100644
index 65a7b3e5519e6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d9a04b7f41dd6f0db017157a44790f35c626e2d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d9c659ba43bb907fd4e3e36a50958288bafd1a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d9c659ba43bb907fd4e3e36a50958288bafd1a3.hip
deleted file mode 100644
index da8986d59ccbe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2d9c659ba43bb907fd4e3e36a50958288bafd1a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2da2b905c4ce32234c2af62328adae6b1f9217a8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2da2b905c4ce32234c2af62328adae6b1f9217a8.hip
deleted file mode 100644
index 93d7b0d6a8265..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2da2b905c4ce32234c2af62328adae6b1f9217a8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2db33b5442d2e0948762b1f2147a321a9d6907be.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2db33b5442d2e0948762b1f2147a321a9d6907be.hip
deleted file mode 100644
index a7f9f5371d006..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2db33b5442d2e0948762b1f2147a321a9d6907be.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2dfac5a83def98340c8786d55a30a98ad68b9eed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2dfac5a83def98340c8786d55a30a98ad68b9eed.hip
deleted file mode 100644
index d9d2e9b7b664a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2dfac5a83def98340c8786d55a30a98ad68b9eed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e30f50071113dc4ab59468d568ac9deb06b0342.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e30f50071113dc4ab59468d568ac9deb06b0342.hip
deleted file mode 100644
index 46be21c690e87..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e30f50071113dc4ab59468d568ac9deb06b0342.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e43e401abbfb1b6737e4dc822f68421abbc648a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e43e401abbfb1b6737e4dc822f68421abbc648a.hip
deleted file mode 100644
index 874de292b43bb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e43e401abbfb1b6737e4dc822f68421abbc648a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e8b4260626beeac76c26dbcee3cba1457b30e99.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e8b4260626beeac76c26dbcee3cba1457b30e99.hip
deleted file mode 100644
index 1efa353334df3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2e8b4260626beeac76c26dbcee3cba1457b30e99.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ea394a09c8691a534ad2219bedf73724b6dd5ce.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ea394a09c8691a534ad2219bedf73724b6dd5ce.hip
deleted file mode 100644
index 005eace593b7b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2ea394a09c8691a534ad2219bedf73724b6dd5ce.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2eba937ff6d0302ab013db7349d4feb914107f1f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2eba937ff6d0302ab013db7349d4feb914107f1f.hip
deleted file mode 100644
index 4a6be87356cf6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2eba937ff6d0302ab013db7349d4feb914107f1f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f0247e301a7b076b6ec8a778c3b47e330638963.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f0247e301a7b076b6ec8a778c3b47e330638963.hip
deleted file mode 100644
index 691f70712fe11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f0247e301a7b076b6ec8a778c3b47e330638963.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f32f2d658f1f69840fbad511ce8a3851c859d52.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f32f2d658f1f69840fbad511ce8a3851c859d52.hip
deleted file mode 100644
index 74323057aa21d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f32f2d658f1f69840fbad511ce8a3851c859d52.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f55a23a0f24ff7062a4c286944f25d2db3e20a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f55a23a0f24ff7062a4c286944f25d2db3e20a4.hip
deleted file mode 100644
index c3a64b0a4f3c0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_2f55a23a0f24ff7062a4c286944f25d2db3e20a4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30024440e780fdf9ec94deccc85216d8bbb5788a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30024440e780fdf9ec94deccc85216d8bbb5788a.hip
deleted file mode 100644
index 4171a7a06c93a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30024440e780fdf9ec94deccc85216d8bbb5788a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_303b7b04496e4db7c1ba2436485dc7c8a4c88448.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_303b7b04496e4db7c1ba2436485dc7c8a4c88448.hip
deleted file mode 100644
index 97d97de0f938b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_303b7b04496e4db7c1ba2436485dc7c8a4c88448.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3076a6de0e2612279e0ed64612f7393856bcc9ac.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3076a6de0e2612279e0ed64612f7393856bcc9ac.hip
deleted file mode 100644
index 2c580f2b05852..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3076a6de0e2612279e0ed64612f7393856bcc9ac.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30c8e4d5c761fda50e010da779e8e4730051d403.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30c8e4d5c761fda50e010da779e8e4730051d403.hip
deleted file mode 100644
index c14d1116af808..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30c8e4d5c761fda50e010da779e8e4730051d403.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30f0200092b0e18d57a9f5e512d565f1c0229436.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30f0200092b0e18d57a9f5e512d565f1c0229436.hip
deleted file mode 100644
index 51f21a30f1690..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_30f0200092b0e18d57a9f5e512d565f1c0229436.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3108502fd29d3a24b32177bcea968121ee809115.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3108502fd29d3a24b32177bcea968121ee809115.hip
deleted file mode 100644
index 4e454ce1be03c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3108502fd29d3a24b32177bcea968121ee809115.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3110540b50e95e99a5cccebe47d9d3a83093c2fb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3110540b50e95e99a5cccebe47d9d3a83093c2fb.hip
deleted file mode 100644
index 668b2420517de..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3110540b50e95e99a5cccebe47d9d3a83093c2fb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_311104394c8bef8d4ecff35c1409221e723a5a8a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_311104394c8bef8d4ecff35c1409221e723a5a8a.hip
deleted file mode 100644
index 51ba80bf06f8c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_311104394c8bef8d4ecff35c1409221e723a5a8a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_311731442b756308c0a869f21b7b8b103aa613e8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_311731442b756308c0a869f21b7b8b103aa613e8.hip
deleted file mode 100644
index 29da0a953166e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_311731442b756308c0a869f21b7b8b103aa613e8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31222e158484773d2257f4a31e3dfbdb68336a8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31222e158484773d2257f4a31e3dfbdb68336a8e.hip
deleted file mode 100644
index 8ee030c03d3a3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31222e158484773d2257f4a31e3dfbdb68336a8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3163272d25bc2db2ffaa1fea87648b45ee68d408.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3163272d25bc2db2ffaa1fea87648b45ee68d408.hip
deleted file mode 100644
index cb431db18d6ea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3163272d25bc2db2ffaa1fea87648b45ee68d408.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_319df310195191895005b30151da8c1afab6c82f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_319df310195191895005b30151da8c1afab6c82f.hip
deleted file mode 100644
index 61c6df30042a0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_319df310195191895005b30151da8c1afab6c82f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31a968898f0bc6366313e41eddb5e3a3ed12dc98.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31a968898f0bc6366313e41eddb5e3a3ed12dc98.hip
deleted file mode 100644
index bd80920d06056..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31a968898f0bc6366313e41eddb5e3a3ed12dc98.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31b807c48c472e9b1311a6037cd98e21d6706889.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31b807c48c472e9b1311a6037cd98e21d6706889.hip
deleted file mode 100644
index a40f5ce8b0c78..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31b807c48c472e9b1311a6037cd98e21d6706889.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31c3760f5978baf9780ce4587ae4c768af0e49d1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31c3760f5978baf9780ce4587ae4c768af0e49d1.hip
deleted file mode 100644
index 544d7460be065..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31c3760f5978baf9780ce4587ae4c768af0e49d1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31c4b866692ba5c3d115482bef4790733863c1fc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31c4b866692ba5c3d115482bef4790733863c1fc.hip
deleted file mode 100644
index 6c35f5dcc1b71..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_31c4b866692ba5c3d115482bef4790733863c1fc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3206cc121ce8955ed59ea3b12b858ee2e0cf82f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3206cc121ce8955ed59ea3b12b858ee2e0cf82f8.hip
deleted file mode 100644
index 9d582e34820ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3206cc121ce8955ed59ea3b12b858ee2e0cf82f8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_320a6196b662a1d3dc7441a9536d825dc356b95d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_320a6196b662a1d3dc7441a9536d825dc356b95d.hip
deleted file mode 100644
index 5545873a7c942..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_320a6196b662a1d3dc7441a9536d825dc356b95d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_321500dd4c41e4d68834814a48a639f5ca36a2fb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_321500dd4c41e4d68834814a48a639f5ca36a2fb.hip
deleted file mode 100644
index d20025a345486..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_321500dd4c41e4d68834814a48a639f5ca36a2fb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_322a86568f89a5a5a165cfffbae9ca6949f2477e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_322a86568f89a5a5a165cfffbae9ca6949f2477e.hip
deleted file mode 100644
index c9b29a04bcf8d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_322a86568f89a5a5a165cfffbae9ca6949f2477e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32438250078ba2a47345ec4955dafb4e4de78a25.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32438250078ba2a47345ec4955dafb4e4de78a25.hip
deleted file mode 100644
index f628d4b2258f1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32438250078ba2a47345ec4955dafb4e4de78a25.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32527660fa7aeb9a951a9f2fc3c53989bd141c48.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32527660fa7aeb9a951a9f2fc3c53989bd141c48.hip
deleted file mode 100644
index f7d57096f4cdb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32527660fa7aeb9a951a9f2fc3c53989bd141c48.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_325fbcb9e503e68fafea08abf86a4951f440850f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_325fbcb9e503e68fafea08abf86a4951f440850f.hip
deleted file mode 100644
index 02396c643c0dc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_325fbcb9e503e68fafea08abf86a4951f440850f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32652a27e8605cef59c8341813b68e7513be23c5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32652a27e8605cef59c8341813b68e7513be23c5.hip
deleted file mode 100644
index b2be8c4138c51..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_32652a27e8605cef59c8341813b68e7513be23c5.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_327e27892bc57f3dec0da24f94f2a483d6c9321b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_327e27892bc57f3dec0da24f94f2a483d6c9321b.hip
deleted file mode 100644
index 1921d88bb6bee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_327e27892bc57f3dec0da24f94f2a483d6c9321b.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_328a311bafd1c153525393b252e4170f8aafb370.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_328a311bafd1c153525393b252e4170f8aafb370.hip
deleted file mode 100644
index 45eda877ecc61..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_328a311bafd1c153525393b252e4170f8aafb370.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33099fcfc218ffdf69edb4f2f0e46121bea9fafc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33099fcfc218ffdf69edb4f2f0e46121bea9fafc.hip
deleted file mode 100644
index 0b8494da39b6e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33099fcfc218ffdf69edb4f2f0e46121bea9fafc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33746071156e9ad46f403a539dc237e0a44122a7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33746071156e9ad46f403a539dc237e0a44122a7.hip
deleted file mode 100644
index 478b0905146bf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33746071156e9ad46f403a539dc237e0a44122a7.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33e7c1e5f41a451c7baff54f7238b220f1bdf8a1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33e7c1e5f41a451c7baff54f7238b220f1bdf8a1.hip
deleted file mode 100644
index 277d1147b742d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_33e7c1e5f41a451c7baff54f7238b220f1bdf8a1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3400f0af03743dce328486f8fc805dd30bd6da31.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3400f0af03743dce328486f8fc805dd30bd6da31.hip
deleted file mode 100644
index 32abdd8b2e490..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3400f0af03743dce328486f8fc805dd30bd6da31.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3408103188e27b3bc55dce0c1716c0b4d32d6494.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3408103188e27b3bc55dce0c1716c0b4d32d6494.hip
deleted file mode 100644
index 3ea3e0e3f944f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3408103188e27b3bc55dce0c1716c0b4d32d6494.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_342d29c85070f488a14b1915f948e5fd69019c99.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_342d29c85070f488a14b1915f948e5fd69019c99.hip
deleted file mode 100644
index f790620e2ab5c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_342d29c85070f488a14b1915f948e5fd69019c99.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_344932e2655d7b32704be8de9a63bbd8c3369f02.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_344932e2655d7b32704be8de9a63bbd8c3369f02.hip
deleted file mode 100644
index 72afe37e6a95a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_344932e2655d7b32704be8de9a63bbd8c3369f02.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_345a939a2491166dc520e9a2b9de7e43671e0c2b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_345a939a2491166dc520e9a2b9de7e43671e0c2b.hip
deleted file mode 100644
index deb452c600780..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_345a939a2491166dc520e9a2b9de7e43671e0c2b.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_345ea796c8d97bfe3b7c9663bf15e2e5e7696235.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_345ea796c8d97bfe3b7c9663bf15e2e5e7696235.hip
deleted file mode 100644
index 8a68e8a8d228e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_345ea796c8d97bfe3b7c9663bf15e2e5e7696235.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_34807a8e90bf1cd839f32fd718afa6469c35a4fa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_34807a8e90bf1cd839f32fd718afa6469c35a4fa.hip
deleted file mode 100644
index 72f31d894cdc5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_34807a8e90bf1cd839f32fd718afa6469c35a4fa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_349241529745bf138552f49d9a93db418663ad65.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_349241529745bf138552f49d9a93db418663ad65.hip
deleted file mode 100644
index 7b3003046a126..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_349241529745bf138552f49d9a93db418663ad65.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_34c2db98d8e2e690f499f41cfd5afb831b756f54.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_34c2db98d8e2e690f499f41cfd5afb831b756f54.hip
deleted file mode 100644
index 964ca3a46d0a0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_34c2db98d8e2e690f499f41cfd5afb831b756f54.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3511c54e6a6f9eec378d8b661121066536195d3a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3511c54e6a6f9eec378d8b661121066536195d3a.hip
deleted file mode 100644
index 37c6017eaed76..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3511c54e6a6f9eec378d8b661121066536195d3a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_351425a006aeeff4d69c8570cb6bf1e1427d2c21.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_351425a006aeeff4d69c8570cb6bf1e1427d2c21.hip
deleted file mode 100644
index fe7c0cb8523fa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_351425a006aeeff4d69c8570cb6bf1e1427d2c21.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_354121d3bad1d448bd413718fa096f54faa12e95.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_354121d3bad1d448bd413718fa096f54faa12e95.hip
deleted file mode 100644
index 199467f7c90b1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_354121d3bad1d448bd413718fa096f54faa12e95.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_356f83cb96d0313abcdb24955edd4264df72aed7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_356f83cb96d0313abcdb24955edd4264df72aed7.hip
deleted file mode 100644
index e424503f28578..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_356f83cb96d0313abcdb24955edd4264df72aed7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_357f7e626135cc9176a295f3d1f336a7c3852688.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_357f7e626135cc9176a295f3d1f336a7c3852688.hip
deleted file mode 100644
index b8f17aa002101..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_357f7e626135cc9176a295f3d1f336a7c3852688.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_358399e756ed5026baf3ab78af17489dc07b9532.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_358399e756ed5026baf3ab78af17489dc07b9532.hip
deleted file mode 100644
index bff4d31f4e6cd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_358399e756ed5026baf3ab78af17489dc07b9532.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_358d28c958c0a831a615a4811d13279b18db09c4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_358d28c958c0a831a615a4811d13279b18db09c4.hip
deleted file mode 100644
index 48de526f8be47..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_358d28c958c0a831a615a4811d13279b18db09c4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3642b78913a853a62dbff8b99d9ae3fa458f461d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3642b78913a853a62dbff8b99d9ae3fa458f461d.hip
deleted file mode 100644
index dc0ebb7be98f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3642b78913a853a62dbff8b99d9ae3fa458f461d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_366662dccf2f650bcd8123c49006c759cd4c0ef6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_366662dccf2f650bcd8123c49006c759cd4c0ef6.hip
deleted file mode 100644
index 92feaaba9d490..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_366662dccf2f650bcd8123c49006c759cd4c0ef6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_367e58867c46d96c9bbaa96eaaa9f93595c9e099.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_367e58867c46d96c9bbaa96eaaa9f93595c9e099.hip
deleted file mode 100644
index f45066e7e6f98..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_367e58867c46d96c9bbaa96eaaa9f93595c9e099.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_36a0a960541bd8a2dc6741579de685b7c0a5f6d7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_36a0a960541bd8a2dc6741579de685b7c0a5f6d7.hip
deleted file mode 100644
index 60a96f173203f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_36a0a960541bd8a2dc6741579de685b7c0a5f6d7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_377b70f54cb2778b5ce3df936b477f775eea8b3c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_377b70f54cb2778b5ce3df936b477f775eea8b3c.hip
deleted file mode 100644
index 06085da0febc7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_377b70f54cb2778b5ce3df936b477f775eea8b3c.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_378759ae25465c32960487375828e23c5f1ac869.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_378759ae25465c32960487375828e23c5f1ac869.hip
deleted file mode 100644
index 41b799af35f4e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_378759ae25465c32960487375828e23c5f1ac869.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_378bf438642e5d863e31145ada2a0688059aa5d9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_378bf438642e5d863e31145ada2a0688059aa5d9.hip
deleted file mode 100644
index 6ef04047e978a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_378bf438642e5d863e31145ada2a0688059aa5d9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_37ad61bf8427a26775969f8a9166fd0bfb7446b4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_37ad61bf8427a26775969f8a9166fd0bfb7446b4.hip
deleted file mode 100644
index 104d78fd220b5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_37ad61bf8427a26775969f8a9166fd0bfb7446b4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_37fe04467e87ec2110f60c7aea0cc9bf2ca07481.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_37fe04467e87ec2110f60c7aea0cc9bf2ca07481.hip
deleted file mode 100644
index c497fbc815085..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_37fe04467e87ec2110f60c7aea0cc9bf2ca07481.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38010c9bf7341588f071f889b7a0b4dcc4e7a14c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38010c9bf7341588f071f889b7a0b4dcc4e7a14c.hip
deleted file mode 100644
index ff242d6a34d32..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38010c9bf7341588f071f889b7a0b4dcc4e7a14c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_381b29d9888365bff0f109d897b508eebfd8a61f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_381b29d9888365bff0f109d897b508eebfd8a61f.hip
deleted file mode 100644
index 1512b864eea07..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_381b29d9888365bff0f109d897b508eebfd8a61f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3824e97d5ecba46e06d5ec1a9456c810d80227a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3824e97d5ecba46e06d5ec1a9456c810d80227a3.hip
deleted file mode 100644
index ac2f0811a97df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3824e97d5ecba46e06d5ec1a9456c810d80227a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38273a2f8e6bbb42ba0b0871b6c95abb34531f33.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38273a2f8e6bbb42ba0b0871b6c95abb34531f33.hip
deleted file mode 100644
index c967125e7ade5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38273a2f8e6bbb42ba0b0871b6c95abb34531f33.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38a5ff72f22e0ad040a281e66b1aca0bf3a2aadb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38a5ff72f22e0ad040a281e66b1aca0bf3a2aadb.hip
deleted file mode 100644
index 753455f974fd2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38a5ff72f22e0ad040a281e66b1aca0bf3a2aadb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38abcbeaa4d33d3150f2b0238bb62ebbfe960980.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38abcbeaa4d33d3150f2b0238bb62ebbfe960980.hip
deleted file mode 100644
index 6d38f8c7d2762..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38abcbeaa4d33d3150f2b0238bb62ebbfe960980.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38b94d76503e13c911781169fbc378517332c42e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38b94d76503e13c911781169fbc378517332c42e.hip
deleted file mode 100644
index 1e5bdfdeecdc9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38b94d76503e13c911781169fbc378517332c42e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38bb367362fe2c4849ded728ec5dd00969ce188f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38bb367362fe2c4849ded728ec5dd00969ce188f.hip
deleted file mode 100644
index bdf3d950c576f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38bb367362fe2c4849ded728ec5dd00969ce188f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38e12dad9e3bafe177ed3c27c833825813e18fc3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38e12dad9e3bafe177ed3c27c833825813e18fc3.hip
deleted file mode 100644
index 4ab3ddfdda93d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38e12dad9e3bafe177ed3c27c833825813e18fc3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38f8a89468cf9c8606cf12a930db062a83cd0ea0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38f8a89468cf9c8606cf12a930db062a83cd0ea0.hip
deleted file mode 100644
index b173f8bb50dd8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_38f8a89468cf9c8606cf12a930db062a83cd0ea0.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3937d9dfb68351de2942e32f35e2ca1ce71edfa8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3937d9dfb68351de2942e32f35e2ca1ce71edfa8.hip
deleted file mode 100644
index 7c37e2cd72104..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3937d9dfb68351de2942e32f35e2ca1ce71edfa8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_39422621a00ff79b2f5ec0dafb957c77693537b3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_39422621a00ff79b2f5ec0dafb957c77693537b3.hip
deleted file mode 100644
index e1815966996c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_39422621a00ff79b2f5ec0dafb957c77693537b3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3967a8807c9451b09227c0f685c18aafeb062fd2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3967a8807c9451b09227c0f685c18aafeb062fd2.hip
deleted file mode 100644
index 227c824a03407..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3967a8807c9451b09227c0f685c18aafeb062fd2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3992d5df4ba2e999caf6889a852db4e1ba078e65.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3992d5df4ba2e999caf6889a852db4e1ba078e65.hip
deleted file mode 100644
index 33029ba1d5e84..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3992d5df4ba2e999caf6889a852db4e1ba078e65.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_39d3071347a0c98f3221104036f477aa13bffa4d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_39d3071347a0c98f3221104036f477aa13bffa4d.hip
deleted file mode 100644
index a2f7bc3cba44a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_39d3071347a0c98f3221104036f477aa13bffa4d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a1dca5feb864e8981387c2d07e62acef1730aa8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a1dca5feb864e8981387c2d07e62acef1730aa8.hip
deleted file mode 100644
index 3b47770a5d3ea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a1dca5feb864e8981387c2d07e62acef1730aa8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a2280997eb6f1d091094fc54cecf42b7c9c3a2d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a2280997eb6f1d091094fc54cecf42b7c9c3a2d.hip
deleted file mode 100644
index 085072ead4c09..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a2280997eb6f1d091094fc54cecf42b7c9c3a2d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a2643099365d0903c799585f41dc1a525ac9f9e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a2643099365d0903c799585f41dc1a525ac9f9e.hip
deleted file mode 100644
index 47ed4c777f348..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a2643099365d0903c799585f41dc1a525ac9f9e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a6b9566559ed2b1c85f2bea1c55e72c41dc47bd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a6b9566559ed2b1c85f2bea1c55e72c41dc47bd.hip
deleted file mode 100644
index b5ed7194c0d36..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3a6b9566559ed2b1c85f2bea1c55e72c41dc47bd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3af86f458fb4dfcceb7db3357fbae0dc15142a15.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3af86f458fb4dfcceb7db3357fbae0dc15142a15.hip
deleted file mode 100644
index 2447c84a72d0b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3af86f458fb4dfcceb7db3357fbae0dc15142a15.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3afbb5ac9048a962a60f48886728220ae6c2aeaf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3afbb5ac9048a962a60f48886728220ae6c2aeaf.hip
deleted file mode 100644
index c356a161aeef4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3afbb5ac9048a962a60f48886728220ae6c2aeaf.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b26eafe76cca8e74e819220b6de1f4279d48e43.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b26eafe76cca8e74e819220b6de1f4279d48e43.hip
deleted file mode 100644
index 3ef4c0344bb44..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b26eafe76cca8e74e819220b6de1f4279d48e43.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b4ecb47f9ebe8c2784976c3e9bbe4834b475cf1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b4ecb47f9ebe8c2784976c3e9bbe4834b475cf1.hip
deleted file mode 100644
index 5231a4a905c4b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b4ecb47f9ebe8c2784976c3e9bbe4834b475cf1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b508b92f7e123b21658f6e17d624ffa87831fee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b508b92f7e123b21658f6e17d624ffa87831fee.hip
deleted file mode 100644
index f4eec82b6d6ef..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b508b92f7e123b21658f6e17d624ffa87831fee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b5b3c218e4a7b459e54080e24c5b730221eac02.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b5b3c218e4a7b459e54080e24c5b730221eac02.hip
deleted file mode 100644
index 88e828ab4e51a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3b5b3c218e4a7b459e54080e24c5b730221eac02.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bb129e6dee6848043dd0e8fa812ae80fec4d014.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bb129e6dee6848043dd0e8fa812ae80fec4d014.hip
deleted file mode 100644
index 5c9a66cff04c4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bb129e6dee6848043dd0e8fa812ae80fec4d014.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bb3b682eab96e4e173affad75b9d8e73f1dd690.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bb3b682eab96e4e173affad75b9d8e73f1dd690.hip
deleted file mode 100644
index 9e9c2703d2501..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bb3b682eab96e4e173affad75b9d8e73f1dd690.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3be7cea6df8e6dd56194e1172f28943667f1c4ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3be7cea6df8e6dd56194e1172f28943667f1c4ef.hip
deleted file mode 100644
index 97f8bdbeba390..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3be7cea6df8e6dd56194e1172f28943667f1c4ef.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bed3aaf24c73073c604a3b23bb4b0358b8e3490.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bed3aaf24c73073c604a3b23bb4b0358b8e3490.hip
deleted file mode 100644
index 739a87ca491c9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3bed3aaf24c73073c604a3b23bb4b0358b8e3490.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c1454ffc1418dac641f63671e947d9f550b1f0c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c1454ffc1418dac641f63671e947d9f550b1f0c.hip
deleted file mode 100644
index 0578fcf163d67..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c1454ffc1418dac641f63671e947d9f550b1f0c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c38bb80e9880335faaea81985ed5d0e713ecb08.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c38bb80e9880335faaea81985ed5d0e713ecb08.hip
deleted file mode 100644
index 98903225cd753..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c38bb80e9880335faaea81985ed5d0e713ecb08.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c3b7e4b8c1efe59f79a15512716fce2282a79a7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c3b7e4b8c1efe59f79a15512716fce2282a79a7.hip
deleted file mode 100644
index f98b69cdec054..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c3b7e4b8c1efe59f79a15512716fce2282a79a7.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c64c33870ebc329921cfa3867d58b1857421f65.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c64c33870ebc329921cfa3867d58b1857421f65.hip
deleted file mode 100644
index fcdfd087f3b4d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3c64c33870ebc329921cfa3867d58b1857421f65.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cb0cee09d633b6f70febbba63a1e090522cfb4a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cb0cee09d633b6f70febbba63a1e090522cfb4a.hip
deleted file mode 100644
index dd315917436f0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cb0cee09d633b6f70febbba63a1e090522cfb4a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cce3baac1e3ca03af0c3f4ee4d0158ad1031e9f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cce3baac1e3ca03af0c3f4ee4d0158ad1031e9f.hip
deleted file mode 100644
index a83180734296f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cce3baac1e3ca03af0c3f4ee4d0158ad1031e9f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3ccf0a9d5a5451da5dbf6075ccea45e4a140550a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3ccf0a9d5a5451da5dbf6075ccea45e4a140550a.hip
deleted file mode 100644
index 9e3620d44cdc7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3ccf0a9d5a5451da5dbf6075ccea45e4a140550a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cd7a9ca49c1149d46f6b05b0fefc41ecaeb6ea1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cd7a9ca49c1149d46f6b05b0fefc41ecaeb6ea1.hip
deleted file mode 100644
index 59adc0e9a6bc7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cd7a9ca49c1149d46f6b05b0fefc41ecaeb6ea1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cf45927b6d931e31e2209685d787efa28eed8ba.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cf45927b6d931e31e2209685d787efa28eed8ba.hip
deleted file mode 100644
index 6d952df6c1550..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3cf45927b6d931e31e2209685d787efa28eed8ba.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d1cea88a2277b87d405025ba256272a1720f88d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d1cea88a2277b87d405025ba256272a1720f88d.hip
deleted file mode 100644
index 1175c6ae5cc16..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d1cea88a2277b87d405025ba256272a1720f88d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d289100991d4c8c362f64c8f6c4ba395c2f3495.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d289100991d4c8c362f64c8f6c4ba395c2f3495.hip
deleted file mode 100644
index 521bb06351595..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d289100991d4c8c362f64c8f6c4ba395c2f3495.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d3f3eb2f5eb1f3287879604892b1c230df85f1d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d3f3eb2f5eb1f3287879604892b1c230df85f1d.hip
deleted file mode 100644
index 6aae42b410cdd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d3f3eb2f5eb1f3287879604892b1c230df85f1d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d45624dc6e33c477c73a155500b015b6c010de8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d45624dc6e33c477c73a155500b015b6c010de8.hip
deleted file mode 100644
index ed04a414d953f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d45624dc6e33c477c73a155500b015b6c010de8.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d55cb42b0096a8ae338ce100f86e378aa1a04c9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d55cb42b0096a8ae338ce100f86e378aa1a04c9.hip
deleted file mode 100644
index 79a154ab3b086..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3d55cb42b0096a8ae338ce100f86e378aa1a04c9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3da8c31f6d5bcaacfa4a21aed4d1d3caecb48922.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3da8c31f6d5bcaacfa4a21aed4d1d3caecb48922.hip
deleted file mode 100644
index 80f9ab1612ada..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3da8c31f6d5bcaacfa4a21aed4d1d3caecb48922.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3dba3cd44f78c950fe7ceaa5f0629dfc607b30f1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3dba3cd44f78c950fe7ceaa5f0629dfc607b30f1.hip
deleted file mode 100644
index 625575ee0b031..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3dba3cd44f78c950fe7ceaa5f0629dfc607b30f1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3dff884e176ec7cff86d17c6afe1ddaa4dd6007d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3dff884e176ec7cff86d17c6afe1ddaa4dd6007d.hip
deleted file mode 100644
index 42332838f992d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3dff884e176ec7cff86d17c6afe1ddaa4dd6007d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e143d88eaa0d9cfea856b2f3a57d1275a656627.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e143d88eaa0d9cfea856b2f3a57d1275a656627.hip
deleted file mode 100644
index ca6dd81fba63c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e143d88eaa0d9cfea856b2f3a57d1275a656627.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e2557f206fd81d82a3b9d59113105040beb891f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e2557f206fd81d82a3b9d59113105040beb891f.hip
deleted file mode 100644
index 6fc591418cc52..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e2557f206fd81d82a3b9d59113105040beb891f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e562e6c3af28b8478020ce3c3bf73c036001c93.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e562e6c3af28b8478020ce3c3bf73c036001c93.hip
deleted file mode 100644
index 0c23ebd4c25bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e562e6c3af28b8478020ce3c3bf73c036001c93.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e61b019e1398a6a3c36143fb84b5ff22c9f4508.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e61b019e1398a6a3c36143fb84b5ff22c9f4508.hip
deleted file mode 100644
index 877e2efbc2d4c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e61b019e1398a6a3c36143fb84b5ff22c9f4508.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e839660557dee9d5bcda9b56940ce23236c5f6d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e839660557dee9d5bcda9b56940ce23236c5f6d.hip
deleted file mode 100644
index 66f78a8f37361..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3e839660557dee9d5bcda9b56940ce23236c5f6d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3eb2ea922daabbba131b90713e06d8caf5f30662.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3eb2ea922daabbba131b90713e06d8caf5f30662.hip
deleted file mode 100644
index a358749f938bd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3eb2ea922daabbba131b90713e06d8caf5f30662.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3ecf565a5a1c4a09887c67ac3b9a019dca427ac0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3ecf565a5a1c4a09887c67ac3b9a019dca427ac0.hip
deleted file mode 100644
index f1e2897cb4a98..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3ecf565a5a1c4a09887c67ac3b9a019dca427ac0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f34433b784d1e405ade3378918641372a30bf6b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f34433b784d1e405ade3378918641372a30bf6b.hip
deleted file mode 100644
index 5c539047c4a33..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f34433b784d1e405ade3378918641372a30bf6b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f5e01b4f2ca8ea10898c39d6570bd74e85f46ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f5e01b4f2ca8ea10898c39d6570bd74e85f46ed.hip
deleted file mode 100644
index fd32ab0e232f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f5e01b4f2ca8ea10898c39d6570bd74e85f46ed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f7315955f555768f24585a50d75e216c40f062d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f7315955f555768f24585a50d75e216c40f062d.hip
deleted file mode 100644
index f8cd7b38bd484..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3f7315955f555768f24585a50d75e216c40f062d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3fad30ff0739ab5dede67a96e859f8c474c245f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3fad30ff0739ab5dede67a96e859f8c474c245f8.hip
deleted file mode 100644
index 2b7997908e84a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3fad30ff0739ab5dede67a96e859f8c474c245f8.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3fcc6893456a559c7d22714116022fc69b372266.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3fcc6893456a559c7d22714116022fc69b372266.hip
deleted file mode 100644
index cde002264828f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_3fcc6893456a559c7d22714116022fc69b372266.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4018b1fcee808b6cccd131418b6ae9e8bf900d8f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4018b1fcee808b6cccd131418b6ae9e8bf900d8f.hip
deleted file mode 100644
index 881592ca3c960..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4018b1fcee808b6cccd131418b6ae9e8bf900d8f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4018f690b6322588041bb467beabd8a7bc79a2e0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4018f690b6322588041bb467beabd8a7bc79a2e0.hip
deleted file mode 100644
index 5f2553e570fe4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4018f690b6322588041bb467beabd8a7bc79a2e0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40357c5e9739eae136a7abf92bc38d3ac94753f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40357c5e9739eae136a7abf92bc38d3ac94753f8.hip
deleted file mode 100644
index ac0f1bb12c2ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40357c5e9739eae136a7abf92bc38d3ac94753f8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4052ca6a3ec02f6559e4bbf1edde42ad2d127c26.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4052ca6a3ec02f6559e4bbf1edde42ad2d127c26.hip
deleted file mode 100644
index e45c558d932f0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4052ca6a3ec02f6559e4bbf1edde42ad2d127c26.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_405e7efa263223148318ae96bd1929b382e994e1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_405e7efa263223148318ae96bd1929b382e994e1.hip
deleted file mode 100644
index 7e0b78f9e1fd5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_405e7efa263223148318ae96bd1929b382e994e1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40aa64439b80ff8dd12498b3e5f6b625da16e285.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40aa64439b80ff8dd12498b3e5f6b625da16e285.hip
deleted file mode 100644
index 827eb5e3c7257..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40aa64439b80ff8dd12498b3e5f6b625da16e285.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40db688a9189e1c47c300d474df946a248a63303.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40db688a9189e1c47c300d474df946a248a63303.hip
deleted file mode 100644
index dbaef536fa3af..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_40db688a9189e1c47c300d474df946a248a63303.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4118e3ab290263ed2576feaf22a1944bf2ddcb7a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4118e3ab290263ed2576feaf22a1944bf2ddcb7a.hip
deleted file mode 100644
index c84ce03b742ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4118e3ab290263ed2576feaf22a1944bf2ddcb7a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_415b183c50dd2663dabe3eb8b780913b778c54ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_415b183c50dd2663dabe3eb8b780913b778c54ab.hip
deleted file mode 100644
index 9eb2f1e6e0b52..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_415b183c50dd2663dabe3eb8b780913b778c54ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4160f6b6d0869740a5a411abd80108f729f810eb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4160f6b6d0869740a5a411abd80108f729f810eb.hip
deleted file mode 100644
index 4f292068029f5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4160f6b6d0869740a5a411abd80108f729f810eb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_417b1cb14b67dc82f614831550f7deb0895bd7e4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_417b1cb14b67dc82f614831550f7deb0895bd7e4.hip
deleted file mode 100644
index e0a7ccf887183..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_417b1cb14b67dc82f614831550f7deb0895bd7e4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_419461cdb5687ebbb7bf0be136071d70420c1619.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_419461cdb5687ebbb7bf0be136071d70420c1619.hip
deleted file mode 100644
index 9d095aefdff42..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_419461cdb5687ebbb7bf0be136071d70420c1619.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_41b68458076e6cb129d3ec793e95b91430a0c8a1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_41b68458076e6cb129d3ec793e95b91430a0c8a1.hip
deleted file mode 100644
index c532a55604844..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_41b68458076e6cb129d3ec793e95b91430a0c8a1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_41db3f29d1940e59dadc357c040ea37a6ff208d9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_41db3f29d1940e59dadc357c040ea37a6ff208d9.hip
deleted file mode 100644
index ab6915fb44e3d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_41db3f29d1940e59dadc357c040ea37a6ff208d9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4217a48a1677bd26cd48e512f1fc8830a8a551b8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4217a48a1677bd26cd48e512f1fc8830a8a551b8.hip
deleted file mode 100644
index 6a778d4d4a715..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4217a48a1677bd26cd48e512f1fc8830a8a551b8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_428ce4e14cf94b284ffa735fe03d923cc74c9fe0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_428ce4e14cf94b284ffa735fe03d923cc74c9fe0.hip
deleted file mode 100644
index fe79b2f2e0668..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_428ce4e14cf94b284ffa735fe03d923cc74c9fe0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_429b82a27571ac91e3631cbdb7e0a58155abf962.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_429b82a27571ac91e3631cbdb7e0a58155abf962.hip
deleted file mode 100644
index 60db1e29bf6d9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_429b82a27571ac91e3631cbdb7e0a58155abf962.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_42e2326066c91452335eac05f25a6311376bd9e5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_42e2326066c91452335eac05f25a6311376bd9e5.hip
deleted file mode 100644
index e54c65f62c8a9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_42e2326066c91452335eac05f25a6311376bd9e5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4306c6c37cf472ad262f53941611b5e60072bdf6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4306c6c37cf472ad262f53941611b5e60072bdf6.hip
deleted file mode 100644
index d6237a5199916..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4306c6c37cf472ad262f53941611b5e60072bdf6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4347e039c003489dd528faf5d710e687321a3fd7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4347e039c003489dd528faf5d710e687321a3fd7.hip
deleted file mode 100644
index 7e988fe4a36b3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4347e039c003489dd528faf5d710e687321a3fd7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4356b3a2ff49f72b91a6b9c215df285f2798ad47.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4356b3a2ff49f72b91a6b9c215df285f2798ad47.hip
deleted file mode 100644
index e8cc8478fc0d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4356b3a2ff49f72b91a6b9c215df285f2798ad47.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4377ac04be3a6cbdbfbe57612a469412812fb5b5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4377ac04be3a6cbdbfbe57612a469412812fb5b5.hip
deleted file mode 100644
index 421c9e3c09315..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4377ac04be3a6cbdbfbe57612a469412812fb5b5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_438e3565f4c720e6c9691b0d33c1392936e2e7ae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_438e3565f4c720e6c9691b0d33c1392936e2e7ae.hip
deleted file mode 100644
index 577d7b2473f4a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_438e3565f4c720e6c9691b0d33c1392936e2e7ae.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4395d3c96b3f4556b9765fd0a3b5701b2fb10948.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4395d3c96b3f4556b9765fd0a3b5701b2fb10948.hip
deleted file mode 100644
index c642a2e79a287..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4395d3c96b3f4556b9765fd0a3b5701b2fb10948.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_43e7c78e8f65be35e2753a0ad5123118555c56b2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_43e7c78e8f65be35e2753a0ad5123118555c56b2.hip
deleted file mode 100644
index f80570bc62096..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_43e7c78e8f65be35e2753a0ad5123118555c56b2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_43f2156a04b18bab55af60e9357f28d8a4604e8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_43f2156a04b18bab55af60e9357f28d8a4604e8e.hip
deleted file mode 100644
index 2a52cb365b4c3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_43f2156a04b18bab55af60e9357f28d8a4604e8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4409f2a7deb027e864afdfc9975d3ab93c5dcc9a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4409f2a7deb027e864afdfc9975d3ab93c5dcc9a.hip
deleted file mode 100644
index b7efdaef3101e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4409f2a7deb027e864afdfc9975d3ab93c5dcc9a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4432c5214c4d40c54ca2d02f0d4785c6d6902370.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4432c5214c4d40c54ca2d02f0d4785c6d6902370.hip
deleted file mode 100644
index c6a83f8b04af6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4432c5214c4d40c54ca2d02f0d4785c6d6902370.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44462715ed5f192532760d6f4c66ff9d4e20e254.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44462715ed5f192532760d6f4c66ff9d4e20e254.hip
deleted file mode 100644
index 6849fbe134fc5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44462715ed5f192532760d6f4c66ff9d4e20e254.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44564dddf8b492d80be54854abb8d1d831e42679.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44564dddf8b492d80be54854abb8d1d831e42679.hip
deleted file mode 100644
index dcdc632209061..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44564dddf8b492d80be54854abb8d1d831e42679.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_445cd8fa559588f4264ce6192f2de3e3065365ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_445cd8fa559588f4264ce6192f2de3e3065365ea.hip
deleted file mode 100644
index b2ec654b46667..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_445cd8fa559588f4264ce6192f2de3e3065365ea.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_445e28a8a51cd435130ded2abc9fc606e522c713.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_445e28a8a51cd435130ded2abc9fc606e522c713.hip
deleted file mode 100644
index 339dd8b7bd3f9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_445e28a8a51cd435130ded2abc9fc606e522c713.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4462b192a64efb60d5484798526278ac7a0fb9fa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4462b192a64efb60d5484798526278ac7a0fb9fa.hip
deleted file mode 100644
index 61c70a8e60daf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4462b192a64efb60d5484798526278ac7a0fb9fa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4466b6c6b2ec3acb40ac1cda432efa1e4e62d9d9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4466b6c6b2ec3acb40ac1cda432efa1e4e62d9d9.hip
deleted file mode 100644
index dff2e53698f23..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4466b6c6b2ec3acb40ac1cda432efa1e4e62d9d9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44690e48f30657b0fcfa26fb3b9af3ef76e792e3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44690e48f30657b0fcfa26fb3b9af3ef76e792e3.hip
deleted file mode 100644
index 23e4a4ab96ac1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44690e48f30657b0fcfa26fb3b9af3ef76e792e3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44c181996532676f2140fd026707135144e9d37b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44c181996532676f2140fd026707135144e9d37b.hip
deleted file mode 100644
index 77ec3c0090b0c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44c181996532676f2140fd026707135144e9d37b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44cc95831c347212021c0bab7b43acd7daabce42.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44cc95831c347212021c0bab7b43acd7daabce42.hip
deleted file mode 100644
index 51848f9693015..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44cc95831c347212021c0bab7b43acd7daabce42.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44d82b58fdc3e5b7a7c20490ce7f5acce4e6ec79.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44d82b58fdc3e5b7a7c20490ce7f5acce4e6ec79.hip
deleted file mode 100644
index 91b73b0dbb774..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_44d82b58fdc3e5b7a7c20490ce7f5acce4e6ec79.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_451fbbdc2dcf2ec81efce34673ee6c425cc16ca2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_451fbbdc2dcf2ec81efce34673ee6c425cc16ca2.hip
deleted file mode 100644
index ec99d02933eb6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_451fbbdc2dcf2ec81efce34673ee6c425cc16ca2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4568af1b2f104664fd05d21ad789aed39ecfa42b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4568af1b2f104664fd05d21ad789aed39ecfa42b.hip
deleted file mode 100644
index 3182b0b458956..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4568af1b2f104664fd05d21ad789aed39ecfa42b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_457eaffbff3c58183a656687010daa2c16cfc26e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_457eaffbff3c58183a656687010daa2c16cfc26e.hip
deleted file mode 100644
index 91674b036d6cd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_457eaffbff3c58183a656687010daa2c16cfc26e.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_458d708d13577f2b92e6d5adfe952a87e0cf7be5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_458d708d13577f2b92e6d5adfe952a87e0cf7be5.hip
deleted file mode 100644
index b50620ae2c27e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_458d708d13577f2b92e6d5adfe952a87e0cf7be5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_459c8fb6028991321b09a990c2188d854d940268.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_459c8fb6028991321b09a990c2188d854d940268.hip
deleted file mode 100644
index 2eeab76404332..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_459c8fb6028991321b09a990c2188d854d940268.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_459ea3713aef9b916e1b38a882a45012930924d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_459ea3713aef9b916e1b38a882a45012930924d3.hip
deleted file mode 100644
index 78376d9504f3b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_459ea3713aef9b916e1b38a882a45012930924d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_45b9871c220c0065d74bffeed4021d0304a9625c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_45b9871c220c0065d74bffeed4021d0304a9625c.hip
deleted file mode 100644
index 2374b95a48bc5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_45b9871c220c0065d74bffeed4021d0304a9625c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_45f4363f50af1e7ccd24751d5f5b181bf32c604f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_45f4363f50af1e7ccd24751d5f5b181bf32c604f.hip
deleted file mode 100644
index 5bc7f4efab035..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_45f4363f50af1e7ccd24751d5f5b181bf32c604f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4601680af41c8738089ff377147e0547dcad114d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4601680af41c8738089ff377147e0547dcad114d.hip
deleted file mode 100644
index 8c4c055b966b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4601680af41c8738089ff377147e0547dcad114d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_461737a13e24009bf1a5a4b780175043a9f2e33e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_461737a13e24009bf1a5a4b780175043a9f2e33e.hip
deleted file mode 100644
index 7f39d79c06472..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_461737a13e24009bf1a5a4b780175043a9f2e33e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4666db0ff7b035e54f2c0e59acedc2131b722a55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4666db0ff7b035e54f2c0e59acedc2131b722a55.hip
deleted file mode 100644
index ea47087bdc0ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4666db0ff7b035e54f2c0e59acedc2131b722a55.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_468a5f057fd5cef2df5f919f5102f47e86901e3b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_468a5f057fd5cef2df5f919f5102f47e86901e3b.hip
deleted file mode 100644
index 57e1edfe630e9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_468a5f057fd5cef2df5f919f5102f47e86901e3b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_474fe2d739eca8c93fdcb2c105d4154cee6ca1c1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_474fe2d739eca8c93fdcb2c105d4154cee6ca1c1.hip
deleted file mode 100644
index 71b4a27dc4c84..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_474fe2d739eca8c93fdcb2c105d4154cee6ca1c1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47548aa042c69bb9c59a8bf706b44028aaa41830.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47548aa042c69bb9c59a8bf706b44028aaa41830.hip
deleted file mode 100644
index d6ce22e239e1a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47548aa042c69bb9c59a8bf706b44028aaa41830.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47f3ced9b5ddb0dfee8ed5e7df8eca0bbe273047.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47f3ced9b5ddb0dfee8ed5e7df8eca0bbe273047.hip
deleted file mode 100644
index 8a4327297bfaa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47f3ced9b5ddb0dfee8ed5e7df8eca0bbe273047.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47fe73f04cef91cd2a0682e905483968ff80eadb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47fe73f04cef91cd2a0682e905483968ff80eadb.hip
deleted file mode 100644
index 078ae2b27172c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_47fe73f04cef91cd2a0682e905483968ff80eadb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_481415463f0316ebe25ff2fda47c68cc54db3359.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_481415463f0316ebe25ff2fda47c68cc54db3359.hip
deleted file mode 100644
index 1346c8a21eff9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_481415463f0316ebe25ff2fda47c68cc54db3359.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4824e1f8cda50f80988857611da766685da94494.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4824e1f8cda50f80988857611da766685da94494.hip
deleted file mode 100644
index f8948c4ee5f46..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4824e1f8cda50f80988857611da766685da94494.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48280c91d7cd8712fd533e246a6b0f758834abc9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48280c91d7cd8712fd533e246a6b0f758834abc9.hip
deleted file mode 100644
index 241555722550e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48280c91d7cd8712fd533e246a6b0f758834abc9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_482e34930d11ff493007b1613993e01acc1af78d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_482e34930d11ff493007b1613993e01acc1af78d.hip
deleted file mode 100644
index 983981295cc87..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_482e34930d11ff493007b1613993e01acc1af78d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48300e0aeabe337785d4c7b41796ce65df6cc42a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48300e0aeabe337785d4c7b41796ce65df6cc42a.hip
deleted file mode 100644
index b09f2f366fe68..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48300e0aeabe337785d4c7b41796ce65df6cc42a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_483eaea4096c8f5bee16a64860432f0634a253d8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_483eaea4096c8f5bee16a64860432f0634a253d8.hip
deleted file mode 100644
index 9274da55895a3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_483eaea4096c8f5bee16a64860432f0634a253d8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48435e5dd23e49e19dd313f9891ffec800ce74c2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48435e5dd23e49e19dd313f9891ffec800ce74c2.hip
deleted file mode 100644
index 65272f5721edf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48435e5dd23e49e19dd313f9891ffec800ce74c2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_486f6c7c7655c34b7b9973ff357b0813f0a3fd7c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_486f6c7c7655c34b7b9973ff357b0813f0a3fd7c.hip
deleted file mode 100644
index 93b2a7bfd4e9b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_486f6c7c7655c34b7b9973ff357b0813f0a3fd7c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_487724686efd35731e5335efa949486c93ae26e3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_487724686efd35731e5335efa949486c93ae26e3.hip
deleted file mode 100644
index 765e398fd8d16..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_487724686efd35731e5335efa949486c93ae26e3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_489e7be0f85656d012a6451b65f6c1d2613b187d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_489e7be0f85656d012a6451b65f6c1d2613b187d.hip
deleted file mode 100644
index 6b4f865839da9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_489e7be0f85656d012a6451b65f6c1d2613b187d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48ae3af78583258c4b13c11a442022e0e058bb85.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48ae3af78583258c4b13c11a442022e0e058bb85.hip
deleted file mode 100644
index 4d17028eaba20..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48ae3af78583258c4b13c11a442022e0e058bb85.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48d7d145f96aa8958a9208d0c8887742a8c834fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48d7d145f96aa8958a9208d0c8887742a8c834fd.hip
deleted file mode 100644
index e236e9745ee30..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48d7d145f96aa8958a9208d0c8887742a8c834fd.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48e9e858abf6f77489f3fadc4ee81edacd26705a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48e9e858abf6f77489f3fadc4ee81edacd26705a.hip
deleted file mode 100644
index 3f250a5d4b4b3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_48e9e858abf6f77489f3fadc4ee81edacd26705a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4904c5910a2d0595b39a3f87652a9d1ef4fcbe80.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4904c5910a2d0595b39a3f87652a9d1ef4fcbe80.hip
deleted file mode 100644
index 567274c16d5b7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4904c5910a2d0595b39a3f87652a9d1ef4fcbe80.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_490a68220a7b621ae9817d7b77f55de239b0a4f3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_490a68220a7b621ae9817d7b77f55de239b0a4f3.hip
deleted file mode 100644
index f007a0f3896c9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_490a68220a7b621ae9817d7b77f55de239b0a4f3.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4911bdd71351610d55916d452495e599960d0a41.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4911bdd71351610d55916d452495e599960d0a41.hip
deleted file mode 100644
index 8810e69fc430d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4911bdd71351610d55916d452495e599960d0a41.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_492fbc418e829f89bcb8d93f8afd2869dd8dfccc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_492fbc418e829f89bcb8d93f8afd2869dd8dfccc.hip
deleted file mode 100644
index ce3c284e469df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_492fbc418e829f89bcb8d93f8afd2869dd8dfccc.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_49d4c005d723cdab9fbc307933c1257d114b539e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_49d4c005d723cdab9fbc307933c1257d114b539e.hip
deleted file mode 100644
index 3d1411a21a885..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_49d4c005d723cdab9fbc307933c1257d114b539e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_49f5017cc0f5c8c8dc71492e7765cf729c1f225c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_49f5017cc0f5c8c8dc71492e7765cf729c1f225c.hip
deleted file mode 100644
index c56084d7dfd8e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_49f5017cc0f5c8c8dc71492e7765cf729c1f225c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a06b5b153ea6e8b1e20d9aad9d4633333fd98f5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a06b5b153ea6e8b1e20d9aad9d4633333fd98f5.hip
deleted file mode 100644
index e56a4f3b1a795..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a06b5b153ea6e8b1e20d9aad9d4633333fd98f5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a2e6b05e7e4de2cb23d815f8b2c8adf22131c0c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a2e6b05e7e4de2cb23d815f8b2c8adf22131c0c.hip
deleted file mode 100644
index 69681c5b6f738..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a2e6b05e7e4de2cb23d815f8b2c8adf22131c0c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a4a00bd6ea27ff20a2903d619e1361b5e27672a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a4a00bd6ea27ff20a2903d619e1361b5e27672a.hip
deleted file mode 100644
index 53bb0bfcea134..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a4a00bd6ea27ff20a2903d619e1361b5e27672a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a5dbf601de5754c03a03a1a42395dc0766fb8ac.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a5dbf601de5754c03a03a1a42395dc0766fb8ac.hip
deleted file mode 100644
index 8452a203c0a5f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a5dbf601de5754c03a03a1a42395dc0766fb8ac.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a9f3da698a6103caf25d785928dd9f814ac27b4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a9f3da698a6103caf25d785928dd9f814ac27b4.hip
deleted file mode 100644
index 5f38459e753c6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4a9f3da698a6103caf25d785928dd9f814ac27b4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ab5d6e8fbfd92e9f7e47bda5cfbb0d4162a6319.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ab5d6e8fbfd92e9f7e47bda5cfbb0d4162a6319.hip
deleted file mode 100644
index 748731a663a7b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ab5d6e8fbfd92e9f7e47bda5cfbb0d4162a6319.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4afd02981f92fbef6277c1985cc479c12bae9239.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4afd02981f92fbef6277c1985cc479c12bae9239.hip
deleted file mode 100644
index 55927650b0509..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4afd02981f92fbef6277c1985cc479c12bae9239.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b1eaca3c37a82d19f8dc91f06764170069ca3af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b1eaca3c37a82d19f8dc91f06764170069ca3af.hip
deleted file mode 100644
index 0a82ef600335a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b1eaca3c37a82d19f8dc91f06764170069ca3af.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b2e7f96b095ebfb66ecc7a75752fba2a63e4f37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b2e7f96b095ebfb66ecc7a75752fba2a63e4f37.hip
deleted file mode 100644
index b69be439eabf8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b2e7f96b095ebfb66ecc7a75752fba2a63e4f37.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b30f472f00bec9da0564ddc40e07112b5f9a117.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b30f472f00bec9da0564ddc40e07112b5f9a117.hip
deleted file mode 100644
index 576ce59aee2df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b30f472f00bec9da0564ddc40e07112b5f9a117.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b45948f2795293e72530b02669c4f549608ea7f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b45948f2795293e72530b02669c4f549608ea7f.hip
deleted file mode 100644
index 8cfb741bd114e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b45948f2795293e72530b02669c4f549608ea7f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b4c03c916393d6be7c5181369ebcef949eaa763.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b4c03c916393d6be7c5181369ebcef949eaa763.hip
deleted file mode 100644
index 2e1dc8be81290..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b4c03c916393d6be7c5181369ebcef949eaa763.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b68e4d00295b294320b94bc777d7d34609127e0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b68e4d00295b294320b94bc777d7d34609127e0.hip
deleted file mode 100644
index 7d0265e222e52..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b68e4d00295b294320b94bc777d7d34609127e0.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b7393d55600c9892558248f4131fc06a6cf3309.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b7393d55600c9892558248f4131fc06a6cf3309.hip
deleted file mode 100644
index 0288fc1a0391c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b7393d55600c9892558248f4131fc06a6cf3309.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b74439f42140cdda9bb0f78d995d741212a35f4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b74439f42140cdda9bb0f78d995d741212a35f4.hip
deleted file mode 100644
index 9cd3049dd7753..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b74439f42140cdda9bb0f78d995d741212a35f4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b76e5dce9af523422782dd25d8dcf6f25edc68f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b76e5dce9af523422782dd25d8dcf6f25edc68f.hip
deleted file mode 100644
index 20165a311b741..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4b76e5dce9af523422782dd25d8dcf6f25edc68f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4baf664bfdf070362bcc91af77d1bc406f744351.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4baf664bfdf070362bcc91af77d1bc406f744351.hip
deleted file mode 100644
index 80570a482eea9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4baf664bfdf070362bcc91af77d1bc406f744351.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bc48576f285325345fa1205e5e7e01787b74f71.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bc48576f285325345fa1205e5e7e01787b74f71.hip
deleted file mode 100644
index 8fff57e47c0b5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bc48576f285325345fa1205e5e7e01787b74f71.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bd4d46397a3749646b232b306688e52b8c6e584.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bd4d46397a3749646b232b306688e52b8c6e584.hip
deleted file mode 100644
index 8f6a8997a80cc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bd4d46397a3749646b232b306688e52b8c6e584.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4be4a98f150f3f9ab6f03b5fd0968c5454565c9a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4be4a98f150f3f9ab6f03b5fd0968c5454565c9a.hip
deleted file mode 100644
index c57d9dc56cde4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4be4a98f150f3f9ab6f03b5fd0968c5454565c9a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4beca56234ff6fb4f23b9b24822887fd9a3d0df9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4beca56234ff6fb4f23b9b24822887fd9a3d0df9.hip
deleted file mode 100644
index ed0665c06de5f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4beca56234ff6fb4f23b9b24822887fd9a3d0df9.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bef4d120e71bfcfe61d67aa44d24ceb907c2b9e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bef4d120e71bfcfe61d67aa44d24ceb907c2b9e.hip
deleted file mode 100644
index ae14e64399500..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4bef4d120e71bfcfe61d67aa44d24ceb907c2b9e.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c0c50a1fac82d47dff2357ee3ddbfa0b2c8d487.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c0c50a1fac82d47dff2357ee3ddbfa0b2c8d487.hip
deleted file mode 100644
index f42d922d6e848..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c0c50a1fac82d47dff2357ee3ddbfa0b2c8d487.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c69d06e3f32e3b6d28d3e54ad764b472741c193.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c69d06e3f32e3b6d28d3e54ad764b472741c193.hip
deleted file mode 100644
index 732661f9dcdef..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c69d06e3f32e3b6d28d3e54ad764b472741c193.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c8720923c3452e3aebd7b9c1b4b23f0c35d7e4f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c8720923c3452e3aebd7b9c1b4b23f0c35d7e4f.hip
deleted file mode 100644
index ef2b8892fd637..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4c8720923c3452e3aebd7b9c1b4b23f0c35d7e4f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cabdafad0bf803223ba5e8f474cd59233dc48cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cabdafad0bf803223ba5e8f474cd59233dc48cb.hip
deleted file mode 100644
index ac3556c259a58..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cabdafad0bf803223ba5e8f474cd59233dc48cb.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cb1861e31df98bdfd731efc3d335055090d83af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cb1861e31df98bdfd731efc3d335055090d83af.hip
deleted file mode 100644
index cc1405c6e9f99..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cb1861e31df98bdfd731efc3d335055090d83af.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cd3de43cc1f7588d62a10362f59d113ee818846.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cd3de43cc1f7588d62a10362f59d113ee818846.hip
deleted file mode 100644
index 3e7751fbfb2e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4cd3de43cc1f7588d62a10362f59d113ee818846.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ce03571f1d2779bdeaf0a6a2d617e236d191c11.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ce03571f1d2779bdeaf0a6a2d617e236d191c11.hip
deleted file mode 100644
index 09a279e493ad5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ce03571f1d2779bdeaf0a6a2d617e236d191c11.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ce671f5defd76ca08614a7a1f184c36c0f1e2ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ce671f5defd76ca08614a7a1f184c36c0f1e2ab.hip
deleted file mode 100644
index d2b3f1a432663..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ce671f5defd76ca08614a7a1f184c36c0f1e2ab.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d3b1ae63e127b6e6afe39e354d4995afc5faeaf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d3b1ae63e127b6e6afe39e354d4995afc5faeaf.hip
deleted file mode 100644
index 426db612bc3c9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d3b1ae63e127b6e6afe39e354d4995afc5faeaf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d5f3cf0f78f73df79665c26b20b0805615e1b04.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d5f3cf0f78f73df79665c26b20b0805615e1b04.hip
deleted file mode 100644
index 286759ba9eed9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d5f3cf0f78f73df79665c26b20b0805615e1b04.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d65e58c9f147498ed04dd51fe1393770603a6d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d65e58c9f147498ed04dd51fe1393770603a6d3.hip
deleted file mode 100644
index ca68c918803a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d65e58c9f147498ed04dd51fe1393770603a6d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d7dc0f356b630179916f8fc2041b7f1402b46df.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d7dc0f356b630179916f8fc2041b7f1402b46df.hip
deleted file mode 100644
index 6e0b8bc9a5e1a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4d7dc0f356b630179916f8fc2041b7f1402b46df.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4da9e9b7277bc90518ab92860bef2097ba96d982.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4da9e9b7277bc90518ab92860bef2097ba96d982.hip
deleted file mode 100644
index 0650d8a40cc39..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4da9e9b7277bc90518ab92860bef2097ba96d982.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4db2e63cfebcf84043f79be0321708cd159c62b9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4db2e63cfebcf84043f79be0321708cd159c62b9.hip
deleted file mode 100644
index 1ef804b81fa41..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4db2e63cfebcf84043f79be0321708cd159c62b9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dbdd9c3f496a27bde68cf86374999ff2dd53505.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dbdd9c3f496a27bde68cf86374999ff2dd53505.hip
deleted file mode 100644
index fb665d175a893..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dbdd9c3f496a27bde68cf86374999ff2dd53505.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dc87b7d385e7b092e4706c464217b004fd8a6a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dc87b7d385e7b092e4706c464217b004fd8a6a4.hip
deleted file mode 100644
index ab0dc635c08b8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dc87b7d385e7b092e4706c464217b004fd8a6a4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dde56efe17f4fd36a11cc959320a5e43f1dc232.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dde56efe17f4fd36a11cc959320a5e43f1dc232.hip
deleted file mode 100644
index 1344cf44f5eca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4dde56efe17f4fd36a11cc959320a5e43f1dc232.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e0a88ccef04e81b8c684b695f7cb4310e448915.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e0a88ccef04e81b8c684b695f7cb4310e448915.hip
deleted file mode 100644
index b27ec784fadd7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e0a88ccef04e81b8c684b695f7cb4310e448915.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e15e4f16de26068cba30ef12fc29332d45e460e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e15e4f16de26068cba30ef12fc29332d45e460e.hip
deleted file mode 100644
index b1c96c2523773..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e15e4f16de26068cba30ef12fc29332d45e460e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e47f8fa40332c6ed12d9971e0b539049a871c34.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e47f8fa40332c6ed12d9971e0b539049a871c34.hip
deleted file mode 100644
index 96038acc90a62..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e47f8fa40332c6ed12d9971e0b539049a871c34.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e760de14b71a41882ec4a2c7362565af36d1a5d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e760de14b71a41882ec4a2c7362565af36d1a5d.hip
deleted file mode 100644
index 387c14523322e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e760de14b71a41882ec4a2c7362565af36d1a5d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e79dce18e49ffe024fe4cd0693ad3399f5edaee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e79dce18e49ffe024fe4cd0693ad3399f5edaee.hip
deleted file mode 100644
index 89eaeee1488e6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e79dce18e49ffe024fe4cd0693ad3399f5edaee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e9a933b916285d9580a76df543cfafc88a536cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e9a933b916285d9580a76df543cfafc88a536cb.hip
deleted file mode 100644
index 40933938adbfa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4e9a933b916285d9580a76df543cfafc88a536cb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ec2075f394acfb14fae7b1ef4304fd9b654ba0d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ec2075f394acfb14fae7b1ef4304fd9b654ba0d.hip
deleted file mode 100644
index 05471d838e426..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ec2075f394acfb14fae7b1ef4304fd9b654ba0d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ed6da5357b67cc28aee4afa9523adaf055c4e32.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ed6da5357b67cc28aee4afa9523adaf055c4e32.hip
deleted file mode 100644
index f32f9f82806d4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ed6da5357b67cc28aee4afa9523adaf055c4e32.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ef35d82ceb4af2e07719c16109c6d72eaedce67.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ef35d82ceb4af2e07719c16109c6d72eaedce67.hip
deleted file mode 100644
index 72f9cb4d3fa25..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ef35d82ceb4af2e07719c16109c6d72eaedce67.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f0aded9d1baec3125ce8e176248cb146ca580fa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f0aded9d1baec3125ce8e176248cb146ca580fa.hip
deleted file mode 100644
index 7d43d59aff23e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f0aded9d1baec3125ce8e176248cb146ca580fa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f1e1c969b57659e7e1367ac9ba10ed5ef5b69a9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f1e1c969b57659e7e1367ac9ba10ed5ef5b69a9.hip
deleted file mode 100644
index 234fd820a79d7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f1e1c969b57659e7e1367ac9ba10ed5ef5b69a9.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f44435491aa68acb3217b0e693232c67641a2db.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f44435491aa68acb3217b0e693232c67641a2db.hip
deleted file mode 100644
index 457e6642d3d39..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f44435491aa68acb3217b0e693232c67641a2db.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f4a5d56721bb1a1332a65882132a8c5763932ec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f4a5d56721bb1a1332a65882132a8c5763932ec.hip
deleted file mode 100644
index 6c62c0f7a9501..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f4a5d56721bb1a1332a65882132a8c5763932ec.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f6243c6850c0a2d2b7bf1476e12f95f187257b6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f6243c6850c0a2d2b7bf1476e12f95f187257b6.hip
deleted file mode 100644
index 12179b6c2b159..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4f6243c6850c0a2d2b7bf1476e12f95f187257b6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fa4d21931b9afcbd70b1567995d3eeb6f9308aa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fa4d21931b9afcbd70b1567995d3eeb6f9308aa.hip
deleted file mode 100644
index 2af9bee76393d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fa4d21931b9afcbd70b1567995d3eeb6f9308aa.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fa883a36a76edb276a66c5d779294f170d6d4b7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fa883a36a76edb276a66c5d779294f170d6d4b7.hip
deleted file mode 100644
index f6fbbbc4bd8ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fa883a36a76edb276a66c5d779294f170d6d4b7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fd34faa8b168e2ac7862641229e6146d3e28aee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fd34faa8b168e2ac7862641229e6146d3e28aee.hip
deleted file mode 100644
index 0740c732d38ca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fd34faa8b168e2ac7862641229e6146d3e28aee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fe530cbf6363a8f08a94728e45e88ecde299e7b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fe530cbf6363a8f08a94728e45e88ecde299e7b.hip
deleted file mode 100644
index 302832389a7b8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4fe530cbf6363a8f08a94728e45e88ecde299e7b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ff20bafbf156fe8fb80bdd84a5d2f3a4a944c1a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ff20bafbf156fe8fb80bdd84a5d2f3a4a944c1a.hip
deleted file mode 100644
index 4c25e48a3586e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_4ff20bafbf156fe8fb80bdd84a5d2f3a4a944c1a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_501dcf3213efd214cc2ce8c9ba0027f991d241b4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_501dcf3213efd214cc2ce8c9ba0027f991d241b4.hip
deleted file mode 100644
index 982d4160835f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_501dcf3213efd214cc2ce8c9ba0027f991d241b4.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5052b2318dbb78b1a82ef03666a35a623f44481b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5052b2318dbb78b1a82ef03666a35a623f44481b.hip
deleted file mode 100644
index ffe1348e13388..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5052b2318dbb78b1a82ef03666a35a623f44481b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5093976cb7b32a8bd28ce92fc13af00a3e21f737.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5093976cb7b32a8bd28ce92fc13af00a3e21f737.hip
deleted file mode 100644
index 309bbeb200b2d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5093976cb7b32a8bd28ce92fc13af00a3e21f737.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50e59bd079f4d205b613056f975fd2b4e372ab10.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50e59bd079f4d205b613056f975fd2b4e372ab10.hip
deleted file mode 100644
index 4e6d5a27f3d37..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50e59bd079f4d205b613056f975fd2b4e372ab10.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50e7b11019fc2299d70869253877319b03388244.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50e7b11019fc2299d70869253877319b03388244.hip
deleted file mode 100644
index 1a0c62228b8f4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50e7b11019fc2299d70869253877319b03388244.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50f887556a3540609649744957651ca667b91774.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50f887556a3540609649744957651ca667b91774.hip
deleted file mode 100644
index 450980c0f2d04..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50f887556a3540609649744957651ca667b91774.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50f915b4d9bd18a3c25a85917392ea4a5e88b349.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50f915b4d9bd18a3c25a85917392ea4a5e88b349.hip
deleted file mode 100644
index dc127fd45ed4e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_50f915b4d9bd18a3c25a85917392ea4a5e88b349.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_515128c6978449b33ce0c35b02a9e9aaad65ef7a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_515128c6978449b33ce0c35b02a9e9aaad65ef7a.hip
deleted file mode 100644
index a1629b2dda38a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_515128c6978449b33ce0c35b02a9e9aaad65ef7a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_522a2a9435103ed405dc1500d31652f1d431a49d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_522a2a9435103ed405dc1500d31652f1d431a49d.hip
deleted file mode 100644
index 7f183308151e3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_522a2a9435103ed405dc1500d31652f1d431a49d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_523e5bf45ec5008aa3aba4773e68a78e122b2fe7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_523e5bf45ec5008aa3aba4773e68a78e122b2fe7.hip
deleted file mode 100644
index e2bbb093a9742..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_523e5bf45ec5008aa3aba4773e68a78e122b2fe7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52688999141a72e61322140db29043ef9f7fbc3d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52688999141a72e61322140db29043ef9f7fbc3d.hip
deleted file mode 100644
index afb5c2530d762..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52688999141a72e61322140db29043ef9f7fbc3d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_526c89b7a04758b4badbf9695b316f877b8bb053.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_526c89b7a04758b4badbf9695b316f877b8bb053.hip
deleted file mode 100644
index 1eecd5f5d21e6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_526c89b7a04758b4badbf9695b316f877b8bb053.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_528db08068589c6e4c096054d26a2e5be63285b6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_528db08068589c6e4c096054d26a2e5be63285b6.hip
deleted file mode 100644
index e996953d78314..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_528db08068589c6e4c096054d26a2e5be63285b6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52a89981a05963efcea7ba5c1e967638beeebbbb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52a89981a05963efcea7ba5c1e967638beeebbbb.hip
deleted file mode 100644
index fe0072f7afe9b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52a89981a05963efcea7ba5c1e967638beeebbbb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52a8a323414448c50571a334f29bc0a38919b61d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52a8a323414448c50571a334f29bc0a38919b61d.hip
deleted file mode 100644
index ee3261f350211..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_52a8a323414448c50571a334f29bc0a38919b61d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_532a6ffd8a21d3e98342fd401f0247f62ca4e038.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_532a6ffd8a21d3e98342fd401f0247f62ca4e038.hip
deleted file mode 100644
index 3aa2b7fd91468..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_532a6ffd8a21d3e98342fd401f0247f62ca4e038.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5344427df3ae9392c4fc4c25c232196828e70648.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5344427df3ae9392c4fc4c25c232196828e70648.hip
deleted file mode 100644
index 5f63c871e0f24..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5344427df3ae9392c4fc4c25c232196828e70648.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5382a30dcf702daae19bd6705864bfe36e09502c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5382a30dcf702daae19bd6705864bfe36e09502c.hip
deleted file mode 100644
index f0758b4c0d7f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5382a30dcf702daae19bd6705864bfe36e09502c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_53bd60bd2afee49b30a583c32a45ae9f2076db08.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_53bd60bd2afee49b30a583c32a45ae9f2076db08.hip
deleted file mode 100644
index 93d2fe6f26e77..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_53bd60bd2afee49b30a583c32a45ae9f2076db08.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5403eec1cdd216d5c4a7ba977e2ef92a0d7fcc8b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5403eec1cdd216d5c4a7ba977e2ef92a0d7fcc8b.hip
deleted file mode 100644
index fa3875103b4fc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5403eec1cdd216d5c4a7ba977e2ef92a0d7fcc8b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_540bd57333c6839ccf5cf2e928edb996bc60c371.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_540bd57333c6839ccf5cf2e928edb996bc60c371.hip
deleted file mode 100644
index 90f0a6920e039..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_540bd57333c6839ccf5cf2e928edb996bc60c371.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_541874a7633e5713720b9d084b6d1c6715a51a17.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_541874a7633e5713720b9d084b6d1c6715a51a17.hip
deleted file mode 100644
index 86e7686be8592..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_541874a7633e5713720b9d084b6d1c6715a51a17.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54208a6e8c5263e38f9ffcb062564ab61d2785ff.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54208a6e8c5263e38f9ffcb062564ab61d2785ff.hip
deleted file mode 100644
index 53243c8a54823..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54208a6e8c5263e38f9ffcb062564ab61d2785ff.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5435b4651a90e331fcdcf224282457e3dc038a30.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5435b4651a90e331fcdcf224282457e3dc038a30.hip
deleted file mode 100644
index 9dd6274ab0deb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5435b4651a90e331fcdcf224282457e3dc038a30.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54402a22ceee3b665a3f24edb98b8398c35c6f5a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54402a22ceee3b665a3f24edb98b8398c35c6f5a.hip
deleted file mode 100644
index 6e28a45bcb61f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54402a22ceee3b665a3f24edb98b8398c35c6f5a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54548ad36fb92d0963893146c8db20f53cbf0c8f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54548ad36fb92d0963893146c8db20f53cbf0c8f.hip
deleted file mode 100644
index b121193ace71a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54548ad36fb92d0963893146c8db20f53cbf0c8f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5467aea26852aa9a9e3dae76b906005ddf6fbae1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5467aea26852aa9a9e3dae76b906005ddf6fbae1.hip
deleted file mode 100644
index a3661f00788c0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5467aea26852aa9a9e3dae76b906005ddf6fbae1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_548b347672451e8391388a400d016803f4c4cf8d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_548b347672451e8391388a400d016803f4c4cf8d.hip
deleted file mode 100644
index 5636d23a05258..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_548b347672451e8391388a400d016803f4c4cf8d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54940ce53998becf9bddf56df7d19894a7658168.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54940ce53998becf9bddf56df7d19894a7658168.hip
deleted file mode 100644
index f1ecf895525e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54940ce53998becf9bddf56df7d19894a7658168.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_549b6956eaf678f7eb901567d1a515eddbedae5f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_549b6956eaf678f7eb901567d1a515eddbedae5f.hip
deleted file mode 100644
index 3dbd87c4a277e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_549b6956eaf678f7eb901567d1a515eddbedae5f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54b6e18b10d529eb6b32d7c19c59eaefc7184376.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54b6e18b10d529eb6b32d7c19c59eaefc7184376.hip
deleted file mode 100644
index 15f5ee90ef56f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54b6e18b10d529eb6b32d7c19c59eaefc7184376.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54ff49018f1c12b9fa31e523ad40b9cc162ba34d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54ff49018f1c12b9fa31e523ad40b9cc162ba34d.hip
deleted file mode 100644
index 6b6225859d91b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_54ff49018f1c12b9fa31e523ad40b9cc162ba34d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_555ba79201a585bc091ccfc326fd24e851d1eecc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_555ba79201a585bc091ccfc326fd24e851d1eecc.hip
deleted file mode 100644
index 8118982e01a0a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_555ba79201a585bc091ccfc326fd24e851d1eecc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_556cd05288e1666f5c67fb87ad02ce660e4c589c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_556cd05288e1666f5c67fb87ad02ce660e4c589c.hip
deleted file mode 100644
index dba90557d9dea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_556cd05288e1666f5c67fb87ad02ce660e4c589c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55b14cf2998a61611d1de2594e926fcdc378999c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55b14cf2998a61611d1de2594e926fcdc378999c.hip
deleted file mode 100644
index 57efa0fcf49b9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55b14cf2998a61611d1de2594e926fcdc378999c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55bd9c4f1b7a0621c67f3e964d946ce22fb2fc80.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55bd9c4f1b7a0621c67f3e964d946ce22fb2fc80.hip
deleted file mode 100644
index 1de9a2ed79a47..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55bd9c4f1b7a0621c67f3e964d946ce22fb2fc80.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55bf8444c1c26b91fd490c7216f4d0f8aa0a1f1a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55bf8444c1c26b91fd490c7216f4d0f8aa0a1f1a.hip
deleted file mode 100644
index a80e523302096..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55bf8444c1c26b91fd490c7216f4d0f8aa0a1f1a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55cda610c235987e13232e828f8d86fa88030560.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55cda610c235987e13232e828f8d86fa88030560.hip
deleted file mode 100644
index d9986a2411c35..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55cda610c235987e13232e828f8d86fa88030560.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55ea83a47c6299fefa4220ed88f7a8e1dd938215.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55ea83a47c6299fefa4220ed88f7a8e1dd938215.hip
deleted file mode 100644
index 09c25a93a0793..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_55ea83a47c6299fefa4220ed88f7a8e1dd938215.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_566b4782793c6526bfce7362efbf6bf069928b2b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_566b4782793c6526bfce7362efbf6bf069928b2b.hip
deleted file mode 100644
index de2ea5cc3e12b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_566b4782793c6526bfce7362efbf6bf069928b2b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_566e26d4969bc6bbe9b092bedab11cddb3360c0f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_566e26d4969bc6bbe9b092bedab11cddb3360c0f.hip
deleted file mode 100644
index 75a9d5b8f657e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_566e26d4969bc6bbe9b092bedab11cddb3360c0f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56964a17f902257aca9d08c736516a2c67d9a0e9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56964a17f902257aca9d08c736516a2c67d9a0e9.hip
deleted file mode 100644
index 6a936b22f4f49..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56964a17f902257aca9d08c736516a2c67d9a0e9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56cc4399c5567a9495f17d54c712cc9e65e57521.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56cc4399c5567a9495f17d54c712cc9e65e57521.hip
deleted file mode 100644
index 9cd9b36869baf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56cc4399c5567a9495f17d54c712cc9e65e57521.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56de9a7dfb1201b56528740e9d8a07b62710fcaf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56de9a7dfb1201b56528740e9d8a07b62710fcaf.hip
deleted file mode 100644
index 6093ec75037b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56de9a7dfb1201b56528740e9d8a07b62710fcaf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56ffe9e21362afe9c3a407c09d5de186954931a6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56ffe9e21362afe9c3a407c09d5de186954931a6.hip
deleted file mode 100644
index 4d99dfe2c629a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_56ffe9e21362afe9c3a407c09d5de186954931a6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5724d91c1fd6290a6cf8d52a3801ac6b921dc7d4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5724d91c1fd6290a6cf8d52a3801ac6b921dc7d4.hip
deleted file mode 100644
index b5c6505d18eaa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5724d91c1fd6290a6cf8d52a3801ac6b921dc7d4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_572e68bd619e118292768f0925ccf92cbfa68415.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_572e68bd619e118292768f0925ccf92cbfa68415.hip
deleted file mode 100644
index 4499ce4ba7cae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_572e68bd619e118292768f0925ccf92cbfa68415.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5732094f5917e9164ee0f973ac6ec47245a69101.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5732094f5917e9164ee0f973ac6ec47245a69101.hip
deleted file mode 100644
index c97cb72598e7f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5732094f5917e9164ee0f973ac6ec47245a69101.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5789f267d34c9961ced63ad07ffea2c6d2911415.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5789f267d34c9961ced63ad07ffea2c6d2911415.hip
deleted file mode 100644
index 641ea7b50b619..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5789f267d34c9961ced63ad07ffea2c6d2911415.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5854f09511778dd1779a839b0b194896070f69ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5854f09511778dd1779a839b0b194896070f69ad.hip
deleted file mode 100644
index f7f7a1237ccbf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5854f09511778dd1779a839b0b194896070f69ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58679919fcd292a2a69543de0db94e2985c9d364.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58679919fcd292a2a69543de0db94e2985c9d364.hip
deleted file mode 100644
index 2a287879a9572..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58679919fcd292a2a69543de0db94e2985c9d364.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58762476c7f2bb05dce92ec22c0acbeb03676746.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58762476c7f2bb05dce92ec22c0acbeb03676746.hip
deleted file mode 100644
index e61443fb63efb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58762476c7f2bb05dce92ec22c0acbeb03676746.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_587fc33d02b1932235b8d152e57559060211d591.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_587fc33d02b1932235b8d152e57559060211d591.hip
deleted file mode 100644
index 03286eeef0057..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_587fc33d02b1932235b8d152e57559060211d591.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58a784fb478ff5b3f1e2da9765a3a777efda92e3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58a784fb478ff5b3f1e2da9765a3a777efda92e3.hip
deleted file mode 100644
index 5a74f2a4e48bb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58a784fb478ff5b3f1e2da9765a3a777efda92e3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58a7ab44bbd9fbc97c7805860d5f6ac81d6ae468.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58a7ab44bbd9fbc97c7805860d5f6ac81d6ae468.hip
deleted file mode 100644
index 98811637c874f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58a7ab44bbd9fbc97c7805860d5f6ac81d6ae468.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58eb2edc7738d8d18ac359691da261ceaaf71788.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58eb2edc7738d8d18ac359691da261ceaaf71788.hip
deleted file mode 100644
index 8b747833a795d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_58eb2edc7738d8d18ac359691da261ceaaf71788.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5919133d2ed892745013b2fc5d503414cf0a4d83.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5919133d2ed892745013b2fc5d503414cf0a4d83.hip
deleted file mode 100644
index e6e48e42ab67d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5919133d2ed892745013b2fc5d503414cf0a4d83.hip
+++ /dev/null
@@ -1,14395 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-#include <iostream>
-
-template <typename dot_do_o_trait_, typename dq_dk_dv_trait_, typename convert_dq_trait_>
-float fmha_bwd_(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    if(s.log_level_ > 0)
-        std::cout << ", " << fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_>() << ", " << fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_>() << ", " << fmha_bwd_convert_dq_get_name_<convert_dq_trait_>() << std::flush;
-    return ck_tile::launch_kernel(s,
-        [=](const ck_tile::stream_config& s_){ fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_>(s_, a); },
-        [=](const ck_tile::stream_config& s_){ fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_>(s_, a); },
-        [=](const ck_tile::stream_config& s_){ fmha_bwd_convert_dq_oneshot_<convert_dq_trait_>(s_, a); }
-    );
-}
-
-float fmha_bwd(fmha_bwd_traits t, fmha_bwd_args a, const ck_tile::stream_config& s){
-    float r = -1;
-    if(t.data_type.compare("fp16") == 0){
-        if (t.hdim_q <= 32 && t.hdim_v <= 32) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-        else if (t.hdim_q <= 64 && t.hdim_v <= 64) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-        else if (t.hdim_q <= 128 && t.hdim_v <= 128) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-        else if (t.hdim_q <= 256 && t.hdim_v <= 256) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::fp16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::fp16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-
-    }
-    else if(t.data_type.compare("bf16") == 0){
-        if (t.hdim_q <= 32 && t.hdim_v <= 32) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 != 0) && (a.hdim_v % 32 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 32 == 0) && (a.hdim_v % 32 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<32, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<32, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-        else if (t.hdim_q <= 64 && t.hdim_v <= 64) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 32 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 != 0) && (a.hdim_v % 64 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 64 == 0) && (a.hdim_v % 64 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<64, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<64, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-        else if (t.hdim_q <= 128 && t.hdim_v <= 128) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 != 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 128 == 0) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 != 0) && (a.hdim_v % 128 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 128 == 0) && (a.hdim_v % 128 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<128, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<128, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-        else if (t.hdim_q <= 256 && t.hdim_v <= 256) {
-            if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 != 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 16 == 0 and a.seqlen_q % 64 != 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == false) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (a.seqlen_q % 64 == 0) && (a.seqlen_k % 64 == 0) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, false, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, false, false, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type == mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<false>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::no_bias) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<false, true,  false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 != 0) && (a.hdim_v % 256 != 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, true>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, true, true, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, true, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == true)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, true>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, true>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-            else if((t.is_group_mode == true) && (t.mask_type != mask_enum::no_mask) && (t.bias_type == bias_enum::alibi) && (t.has_dbias == false) && (t.has_dropout == true && t.is_store_randval == false) &&
-                        (true) && (true) && (a.hdim_q % 256 == 0) && (a.hdim_v % 256 == 0) && (t.is_deterministic == false)) {
-                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, true, true, false>;
-                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<256, ck_tile::bf16_t, true, ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP, ck_tile::SimplifiedGenericAttentionMask<true>, ck_tile::BlockDropoutBwd<true,  false, false>, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, true, false, false, false>;
-                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<256, ck_tile::bf16_t, true, true, false, false>;
-                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
-                return r;
-            }
-
-        }
-
-    }
-
-    return r;
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5939e6610e41aff8d1ccdb66d9e84d3e48e8d379.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5939e6610e41aff8d1ccdb66d9e84d3e48e8d379.hip
deleted file mode 100644
index 73d873d839fcf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5939e6610e41aff8d1ccdb66d9e84d3e48e8d379.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_594929c433b049a8cf949ff476309a8faf5c25fb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_594929c433b049a8cf949ff476309a8faf5c25fb.hip
deleted file mode 100644
index f16363e302280..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_594929c433b049a8cf949ff476309a8faf5c25fb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_597a0276ec419f18f060a5186e6bb703ae434ac8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_597a0276ec419f18f060a5186e6bb703ae434ac8.hip
deleted file mode 100644
index 86a32dbb5a02e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_597a0276ec419f18f060a5186e6bb703ae434ac8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59901147b7188212b8d8feea15831a11425fe4b3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59901147b7188212b8d8feea15831a11425fe4b3.hip
deleted file mode 100644
index 6775ca1db2285..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59901147b7188212b8d8feea15831a11425fe4b3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59beb9cb4e161f9dcff79080149076488d436301.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59beb9cb4e161f9dcff79080149076488d436301.hip
deleted file mode 100644
index 182e1c39d90e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59beb9cb4e161f9dcff79080149076488d436301.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59d366421e0b51c90fa53c366d47ed8d51b3a329.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59d366421e0b51c90fa53c366d47ed8d51b3a329.hip
deleted file mode 100644
index e26d9f8407685..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_59d366421e0b51c90fa53c366d47ed8d51b3a329.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a05b4e7782bd0e29ca9f6d33fc59d4304136d41.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a05b4e7782bd0e29ca9f6d33fc59d4304136d41.hip
deleted file mode 100644
index 12d2e95d12f5d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a05b4e7782bd0e29ca9f6d33fc59d4304136d41.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a216f777feec4752f5882677b18168225da4b53.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a216f777feec4752f5882677b18168225da4b53.hip
deleted file mode 100644
index ecf6c4e54e001..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a216f777feec4752f5882677b18168225da4b53.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a29b93cee012c79d4364502f1d90f947c73641d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a29b93cee012c79d4364502f1d90f947c73641d.hip
deleted file mode 100644
index b46adb31b3bd1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a29b93cee012c79d4364502f1d90f947c73641d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a85ae0a16e4b293b549bcb6a3ee52df7fccca32.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a85ae0a16e4b293b549bcb6a3ee52df7fccca32.hip
deleted file mode 100644
index f8cb44c15e5fa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5a85ae0a16e4b293b549bcb6a3ee52df7fccca32.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5aba1183efe205af38e79a1b2dccea5fa515d02e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5aba1183efe205af38e79a1b2dccea5fa515d02e.hip
deleted file mode 100644
index b912bf7fa549f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5aba1183efe205af38e79a1b2dccea5fa515d02e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ace1c9b00f160a17355d4583d49c47887ac33c8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ace1c9b00f160a17355d4583d49c47887ac33c8.hip
deleted file mode 100644
index cfa07b8966f79..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ace1c9b00f160a17355d4583d49c47887ac33c8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5af96b404feac271dac8f4190180754480d3ba80.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5af96b404feac271dac8f4190180754480d3ba80.hip
deleted file mode 100644
index 379c0917decfb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5af96b404feac271dac8f4190180754480d3ba80.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b413bdc825ae863d53dab548f2145dc0de8fd37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b413bdc825ae863d53dab548f2145dc0de8fd37.hip
deleted file mode 100644
index 3ce072783e28f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b413bdc825ae863d53dab548f2145dc0de8fd37.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b55946ff3c15a44b9c741e9f6bbbcb5bd4c8577.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b55946ff3c15a44b9c741e9f6bbbcb5bd4c8577.hip
deleted file mode 100644
index 60b33e6de23e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b55946ff3c15a44b9c741e9f6bbbcb5bd4c8577.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b7a4ea3bb8905a22ae97a94c354b1cbe38093bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b7a4ea3bb8905a22ae97a94c354b1cbe38093bb.hip
deleted file mode 100644
index 07aac70d3566d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5b7a4ea3bb8905a22ae97a94c354b1cbe38093bb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ba578c0e7abf1127dd0370f06d7278656c93ab9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ba578c0e7abf1127dd0370f06d7278656c93ab9.hip
deleted file mode 100644
index c69568dada240..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ba578c0e7abf1127dd0370f06d7278656c93ab9.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5bc803342862aa30e23e5be7d84e611bc571c529.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5bc803342862aa30e23e5be7d84e611bc571c529.hip
deleted file mode 100644
index fc9ffd7c9b56b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5bc803342862aa30e23e5be7d84e611bc571c529.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5be9ed84ad9be1627db7a66af9370679816c0897.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5be9ed84ad9be1627db7a66af9370679816c0897.hip
deleted file mode 100644
index ebf8748269469..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5be9ed84ad9be1627db7a66af9370679816c0897.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5bead6be6e39ece0e5d44335083336f7f546d2f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5bead6be6e39ece0e5d44335083336f7f546d2f8.hip
deleted file mode 100644
index 9ab599a95e0e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5bead6be6e39ece0e5d44335083336f7f546d2f8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5c36fc744dfb0d985c9113175e76c7ec1c935054.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5c36fc744dfb0d985c9113175e76c7ec1c935054.hip
deleted file mode 100644
index 949ebb0054868..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5c36fc744dfb0d985c9113175e76c7ec1c935054.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5c742b9ac6749f189d597ac97d46d35189472c50.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5c742b9ac6749f189d597ac97d46d35189472c50.hip
deleted file mode 100644
index 1ede6dd6ad3b8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5c742b9ac6749f189d597ac97d46d35189472c50.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5cd03e29403ad53d6d52e5e81182ea6ff5aff2be.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5cd03e29403ad53d6d52e5e81182ea6ff5aff2be.hip
deleted file mode 100644
index b2babf12db62b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5cd03e29403ad53d6d52e5e81182ea6ff5aff2be.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5cd41b6f578f3c903eb9d58ebfab62eb296044e0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5cd41b6f578f3c903eb9d58ebfab62eb296044e0.hip
deleted file mode 100644
index 9ce2621ee4779..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5cd41b6f578f3c903eb9d58ebfab62eb296044e0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5d707d065ae152450f9def619ddc3dddb9089e88.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5d707d065ae152450f9def619ddc3dddb9089e88.hip
deleted file mode 100644
index 37d96cb19303b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5d707d065ae152450f9def619ddc3dddb9089e88.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5d7ed4c885fb32a0b548186e56d64bab98071d30.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5d7ed4c885fb32a0b548186e56d64bab98071d30.hip
deleted file mode 100644
index 3dcfaca1089ae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5d7ed4c885fb32a0b548186e56d64bab98071d30.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5daedab8931f2eefb649b91e80145cb71b63360c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5daedab8931f2eefb649b91e80145cb71b63360c.hip
deleted file mode 100644
index 264123892a53a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5daedab8931f2eefb649b91e80145cb71b63360c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5de27c4081377f59363c2bf2ea8624217566d2d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5de27c4081377f59363c2bf2ea8624217566d2d3.hip
deleted file mode 100644
index 2952198108679..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5de27c4081377f59363c2bf2ea8624217566d2d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e0abf4e2b6be3e2c555c2134705b9dcaee617ce.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e0abf4e2b6be3e2c555c2134705b9dcaee617ce.hip
deleted file mode 100644
index 3933d8ee07fa6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e0abf4e2b6be3e2c555c2134705b9dcaee617ce.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e62968de58d9df7d687d671f37d63393f189321.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e62968de58d9df7d687d671f37d63393f189321.hip
deleted file mode 100644
index 1b6a404962239..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e62968de58d9df7d687d671f37d63393f189321.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e735b12d130ebf849ac5d6752e413ecf3e69fbf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e735b12d130ebf849ac5d6752e413ecf3e69fbf.hip
deleted file mode 100644
index 1579ca0b58a48..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e735b12d130ebf849ac5d6752e413ecf3e69fbf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e840be0741afa4d41fd4789c8300223fdc63ddc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e840be0741afa4d41fd4789c8300223fdc63ddc.hip
deleted file mode 100644
index 9b9d4796ee6ae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5e840be0741afa4d41fd4789c8300223fdc63ddc.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ea53f7c6370845fa94aa9b395c52fd1900b62de.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ea53f7c6370845fa94aa9b395c52fd1900b62de.hip
deleted file mode 100644
index 8bf88c68fbf92..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5ea53f7c6370845fa94aa9b395c52fd1900b62de.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5efe77ca5c394a60af0313072cdd132216a52bf3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5efe77ca5c394a60af0313072cdd132216a52bf3.hip
deleted file mode 100644
index 965e6030c662e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5efe77ca5c394a60af0313072cdd132216a52bf3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f20263fd84776f155519b3481be5e2c5b035585.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f20263fd84776f155519b3481be5e2c5b035585.hip
deleted file mode 100644
index 37fca282cd7d6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f20263fd84776f155519b3481be5e2c5b035585.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f3c3bed2b584ea2031debf9f953f5f8f7012171.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f3c3bed2b584ea2031debf9f953f5f8f7012171.hip
deleted file mode 100644
index 55ff1b31f4493..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f3c3bed2b584ea2031debf9f953f5f8f7012171.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f71e663978dbcba859c5114ec675a712e343fd6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f71e663978dbcba859c5114ec675a712e343fd6.hip
deleted file mode 100644
index c7b60899dd412..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f71e663978dbcba859c5114ec675a712e343fd6.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f8925f929a5b26f3544ca31938aa75b3c59d34d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f8925f929a5b26f3544ca31938aa75b3c59d34d.hip
deleted file mode 100644
index 9be7a392e3562..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f8925f929a5b26f3544ca31938aa75b3c59d34d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f954a393b7b5a7131c13d0c4578443f468a738d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f954a393b7b5a7131c13d0c4578443f468a738d.hip
deleted file mode 100644
index 8ea46759a796f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5f954a393b7b5a7131c13d0c4578443f468a738d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fa19223cf296d7fd10e15e2571e63c84a80fbb1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fa19223cf296d7fd10e15e2571e63c84a80fbb1.hip
deleted file mode 100644
index cd1bd1a76308f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fa19223cf296d7fd10e15e2571e63c84a80fbb1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fa7fafd4227918e0c7f0c6ca3b2bd673cd07279.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fa7fafd4227918e0c7f0c6ca3b2bd673cd07279.hip
deleted file mode 100644
index 506f74bc003ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fa7fafd4227918e0c7f0c6ca3b2bd673cd07279.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fb062527121e627871b3f1b2a94b96c42e51205.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fb062527121e627871b3f1b2a94b96c42e51205.hip
deleted file mode 100644
index c69d066678539..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fb062527121e627871b3f1b2a94b96c42e51205.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fc66c5b53f83bf1e023e81e9d51f0285b3ae731.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fc66c5b53f83bf1e023e81e9d51f0285b3ae731.hip
deleted file mode 100644
index 8d679b5a7a1bd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_5fc66c5b53f83bf1e023e81e9d51f0285b3ae731.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6018ab272d7306689c7dc5a6d5326efea1471235.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6018ab272d7306689c7dc5a6d5326efea1471235.hip
deleted file mode 100644
index 07046a15a25ee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6018ab272d7306689c7dc5a6d5326efea1471235.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6049c01db99fce654e9351e711b113cf7424550a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6049c01db99fce654e9351e711b113cf7424550a.hip
deleted file mode 100644
index 16ed12be54878..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6049c01db99fce654e9351e711b113cf7424550a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_606f5e0b99814b0a82a731de36f28024bc317801.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_606f5e0b99814b0a82a731de36f28024bc317801.hip
deleted file mode 100644
index 67d6aa71259b6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_606f5e0b99814b0a82a731de36f28024bc317801.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_60801d21c14796c08377349ec86a6c800af497b7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_60801d21c14796c08377349ec86a6c800af497b7.hip
deleted file mode 100644
index dbe5a640a83ac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_60801d21c14796c08377349ec86a6c800af497b7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6082d55544b5280b49b071ea277fb1827193fa2a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6082d55544b5280b49b071ea277fb1827193fa2a.hip
deleted file mode 100644
index 65998fe6898bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6082d55544b5280b49b071ea277fb1827193fa2a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_609616f72bf16a060fa50091ac139ddc06bf9d88.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_609616f72bf16a060fa50091ac139ddc06bf9d88.hip
deleted file mode 100644
index 2f80789150a9a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_609616f72bf16a060fa50091ac139ddc06bf9d88.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_609f68180582384ba81aae2b1d4a4c52dde2c68c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_609f68180582384ba81aae2b1d4a4c52dde2c68c.hip
deleted file mode 100644
index bf022b31497e8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_609f68180582384ba81aae2b1d4a4c52dde2c68c.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_60efa9c427dc278c0d1bc31189f683cd45e4d873.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_60efa9c427dc278c0d1bc31189f683cd45e4d873.hip
deleted file mode 100644
index cf1c43eead356..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_60efa9c427dc278c0d1bc31189f683cd45e4d873.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61204f6805d5d830aa6fca2a9b5f238ed63c3a73.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61204f6805d5d830aa6fca2a9b5f238ed63c3a73.hip
deleted file mode 100644
index 95260742fc3c5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61204f6805d5d830aa6fca2a9b5f238ed63c3a73.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61220f6dca850a5b5ccf1f619a267c40c37efeca.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61220f6dca850a5b5ccf1f619a267c40c37efeca.hip
deleted file mode 100644
index 7ba98f65bb631..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61220f6dca850a5b5ccf1f619a267c40c37efeca.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_614a9f10ebc51bde3f580ef527c17f89489c12c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_614a9f10ebc51bde3f580ef527c17f89489c12c7.hip
deleted file mode 100644
index 84b4dc89da195..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_614a9f10ebc51bde3f580ef527c17f89489c12c7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_615430cb65d8d540836c7f12b3367abd3c8e63d2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_615430cb65d8d540836c7f12b3367abd3c8e63d2.hip
deleted file mode 100644
index 3c15d4330431d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_615430cb65d8d540836c7f12b3367abd3c8e63d2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_618031345ea71cc17e458eb97a559b7c94d3ae43.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_618031345ea71cc17e458eb97a559b7c94d3ae43.hip
deleted file mode 100644
index 3cb600ee4026a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_618031345ea71cc17e458eb97a559b7c94d3ae43.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61896aa9e4e4d7e494c1755b1e77a08e0e264f8d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61896aa9e4e4d7e494c1755b1e77a08e0e264f8d.hip
deleted file mode 100644
index 7bf0a0cee582a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61896aa9e4e4d7e494c1755b1e77a08e0e264f8d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61a44ac409e914c12281f1d26e5b52d8bfd0df75.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61a44ac409e914c12281f1d26e5b52d8bfd0df75.hip
deleted file mode 100644
index 46d578606ff5a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61a44ac409e914c12281f1d26e5b52d8bfd0df75.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61a9e92183ba87924e73ff0b5e25bd12d6038e69.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61a9e92183ba87924e73ff0b5e25bd12d6038e69.hip
deleted file mode 100644
index b9599a8a5869a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_61a9e92183ba87924e73ff0b5e25bd12d6038e69.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62048a8ae1c0096f3372b0114c15edbe813425fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62048a8ae1c0096f3372b0114c15edbe813425fd.hip
deleted file mode 100644
index 1eaa08a2a3e63..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62048a8ae1c0096f3372b0114c15edbe813425fd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6214f820b39a8ba81e547a78ed19a909ac13221c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6214f820b39a8ba81e547a78ed19a909ac13221c.hip
deleted file mode 100644
index 1b4197acb3ac0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6214f820b39a8ba81e547a78ed19a909ac13221c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_621da34ee666903307d3a09b7a032f2a70054759.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_621da34ee666903307d3a09b7a032f2a70054759.hip
deleted file mode 100644
index 9f641c25b044c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_621da34ee666903307d3a09b7a032f2a70054759.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_628b28f65f19e7d1b22fb3b85b7cf3d09cd54ebc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_628b28f65f19e7d1b22fb3b85b7cf3d09cd54ebc.hip
deleted file mode 100644
index fd55dcf45642f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_628b28f65f19e7d1b22fb3b85b7cf3d09cd54ebc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_629e0b97b3fece7c12504f4c8f1860d611b57269.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_629e0b97b3fece7c12504f4c8f1860d611b57269.hip
deleted file mode 100644
index 242df26fd3c70..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_629e0b97b3fece7c12504f4c8f1860d611b57269.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62ab710e4acc711430745e05e036dd6a4d6bcdca.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62ab710e4acc711430745e05e036dd6a4d6bcdca.hip
deleted file mode 100644
index 05bdb5633063b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62ab710e4acc711430745e05e036dd6a4d6bcdca.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62ba7a5a0f3a714eb5f9f2af20f7bfbc82a30350.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62ba7a5a0f3a714eb5f9f2af20f7bfbc82a30350.hip
deleted file mode 100644
index 615213a2ee39e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62ba7a5a0f3a714eb5f9f2af20f7bfbc82a30350.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62eb2f81e73d65fddce7ff43c397da6529317607.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62eb2f81e73d65fddce7ff43c397da6529317607.hip
deleted file mode 100644
index e3790b15bb31e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_62eb2f81e73d65fddce7ff43c397da6529317607.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_634d530731c7ade2c7beecfd1bbbca8583032217.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_634d530731c7ade2c7beecfd1bbbca8583032217.hip
deleted file mode 100644
index 94a5af87cd131..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_634d530731c7ade2c7beecfd1bbbca8583032217.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6360621af3f7e1e81a8be48fea8d2750fdecbbf4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6360621af3f7e1e81a8be48fea8d2750fdecbbf4.hip
deleted file mode 100644
index c0f0585f00854..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6360621af3f7e1e81a8be48fea8d2750fdecbbf4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6376eb68c550b50b9aea42a7a2cc3bda186b0e40.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6376eb68c550b50b9aea42a7a2cc3bda186b0e40.hip
deleted file mode 100644
index c01596dd82b54..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6376eb68c550b50b9aea42a7a2cc3bda186b0e40.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_63c411351ec59bdbed2590c599f9eddf7807b371.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_63c411351ec59bdbed2590c599f9eddf7807b371.hip
deleted file mode 100644
index 962d0ef7dd572..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_63c411351ec59bdbed2590c599f9eddf7807b371.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_63f121a3c8928c10a2d86b487cd13fa995da670d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_63f121a3c8928c10a2d86b487cd13fa995da670d.hip
deleted file mode 100644
index 4072a55be331d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_63f121a3c8928c10a2d86b487cd13fa995da670d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_643b3798f11997d33ccb58d90ed6c10d5411b735.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_643b3798f11997d33ccb58d90ed6c10d5411b735.hip
deleted file mode 100644
index 7723199302e71..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_643b3798f11997d33ccb58d90ed6c10d5411b735.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_649336d59a8b35919e593217b6fd4314a04ea359.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_649336d59a8b35919e593217b6fd4314a04ea359.hip
deleted file mode 100644
index 36e3e7e23a897..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_649336d59a8b35919e593217b6fd4314a04ea359.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64a0ca185449a49fa485892fde6af745ba758167.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64a0ca185449a49fa485892fde6af745ba758167.hip
deleted file mode 100644
index 252b2a476d3af..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64a0ca185449a49fa485892fde6af745ba758167.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64b3488ddf3bb1a4870371882f0a5d267bdfdf73.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64b3488ddf3bb1a4870371882f0a5d267bdfdf73.hip
deleted file mode 100644
index a85575d7ed583..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64b3488ddf3bb1a4870371882f0a5d267bdfdf73.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64c3c1e3dac623f07c2dc1b934ccb868cafcb38c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64c3c1e3dac623f07c2dc1b934ccb868cafcb38c.hip
deleted file mode 100644
index 35f9729777d96..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64c3c1e3dac623f07c2dc1b934ccb868cafcb38c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64cf03c0aa3f1b2a7b76b4e3418eb5063b982a29.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64cf03c0aa3f1b2a7b76b4e3418eb5063b982a29.hip
deleted file mode 100644
index 0c04e1c6d8540..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64cf03c0aa3f1b2a7b76b4e3418eb5063b982a29.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64fe2db75cb20428856b02cd1cc8d7b393a6ad9c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64fe2db75cb20428856b02cd1cc8d7b393a6ad9c.hip
deleted file mode 100644
index ffc7421e3c228..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_64fe2db75cb20428856b02cd1cc8d7b393a6ad9c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_65794d9c185b21f59274ac5d4db10a7abc0be968.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_65794d9c185b21f59274ac5d4db10a7abc0be968.hip
deleted file mode 100644
index 81ca0165e7d88..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_65794d9c185b21f59274ac5d4db10a7abc0be968.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_658552954505a2092662071401e135e84956c4c0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_658552954505a2092662071401e135e84956c4c0.hip
deleted file mode 100644
index b02514fd7dd11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_658552954505a2092662071401e135e84956c4c0.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_65910c8b7a30acc731948ab58467fdbe4fe32f6d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_65910c8b7a30acc731948ab58467fdbe4fe32f6d.hip
deleted file mode 100644
index 7fcb1c9f007f5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_65910c8b7a30acc731948ab58467fdbe4fe32f6d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_661b49505cfecbe4ec3e5c7371de3aaaa85ac9d5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_661b49505cfecbe4ec3e5c7371de3aaaa85ac9d5.hip
deleted file mode 100644
index 74a4bb85dad7d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_661b49505cfecbe4ec3e5c7371de3aaaa85ac9d5.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_661ffaf653085dd7f122d603bb3ba4b001e5f3c0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_661ffaf653085dd7f122d603bb3ba4b001e5f3c0.hip
deleted file mode 100644
index 13028377a8f03..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_661ffaf653085dd7f122d603bb3ba4b001e5f3c0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_662767e588220d0dc6137b00cc1d8dcc91e97134.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_662767e588220d0dc6137b00cc1d8dcc91e97134.hip
deleted file mode 100644
index 0493f71435c04..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_662767e588220d0dc6137b00cc1d8dcc91e97134.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6649f19deeaea20663bee781af7edced7f7a4fc0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6649f19deeaea20663bee781af7edced7f7a4fc0.hip
deleted file mode 100644
index cc15c86524621..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6649f19deeaea20663bee781af7edced7f7a4fc0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66968bbf7e210911fcb95ba90c79837230ab1ce3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66968bbf7e210911fcb95ba90c79837230ab1ce3.hip
deleted file mode 100644
index 38bd1c2abfa25..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66968bbf7e210911fcb95ba90c79837230ab1ce3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66a020f728df204ff51e37d2ddc21afb0aad5e7b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66a020f728df204ff51e37d2ddc21afb0aad5e7b.hip
deleted file mode 100644
index 59293e5666f94..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66a020f728df204ff51e37d2ddc21afb0aad5e7b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66be70b088b20fc8de464167c35745461ddab640.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66be70b088b20fc8de464167c35745461ddab640.hip
deleted file mode 100644
index 9040582b39f93..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66be70b088b20fc8de464167c35745461ddab640.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66f651d3415562206c1049b172261fddba01ea6c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66f651d3415562206c1049b172261fddba01ea6c.hip
deleted file mode 100644
index 6426bba668341..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_66f651d3415562206c1049b172261fddba01ea6c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_671828f15eec2a58be23063a1a8132d337cd26de.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_671828f15eec2a58be23063a1a8132d337cd26de.hip
deleted file mode 100644
index 6ef9fdb78fb29..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_671828f15eec2a58be23063a1a8132d337cd26de.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6767cce35ab784aa42ebcb75af7305bc38a8721a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6767cce35ab784aa42ebcb75af7305bc38a8721a.hip
deleted file mode 100644
index 7931294f2ed5e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6767cce35ab784aa42ebcb75af7305bc38a8721a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6785dcec0197fdbb50124ab06efa627f1a2c0567.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6785dcec0197fdbb50124ab06efa627f1a2c0567.hip
deleted file mode 100644
index 502e487f780d3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6785dcec0197fdbb50124ab06efa627f1a2c0567.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_678a4a8210a972bb2ed89d6ac754fb79438ab2da.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_678a4a8210a972bb2ed89d6ac754fb79438ab2da.hip
deleted file mode 100644
index 85a609d3e40cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_678a4a8210a972bb2ed89d6ac754fb79438ab2da.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_67fb736c61088b8dd92fe0371f5c98e23bf9077f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_67fb736c61088b8dd92fe0371f5c98e23bf9077f.hip
deleted file mode 100644
index 2bd2319988170..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_67fb736c61088b8dd92fe0371f5c98e23bf9077f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_680e81c3700f130df142c9a37a368944ca548721.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_680e81c3700f130df142c9a37a368944ca548721.hip
deleted file mode 100644
index 675513aee0b66..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_680e81c3700f130df142c9a37a368944ca548721.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_683e8a33fdb7053760c9c135002b0a94facbe015.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_683e8a33fdb7053760c9c135002b0a94facbe015.hip
deleted file mode 100644
index 08a203a63e56f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_683e8a33fdb7053760c9c135002b0a94facbe015.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_687f4aaafd1a5b9ee85aadc6fab79ad0c27a2ea2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_687f4aaafd1a5b9ee85aadc6fab79ad0c27a2ea2.hip
deleted file mode 100644
index c14f7ad41ed10..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_687f4aaafd1a5b9ee85aadc6fab79ad0c27a2ea2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_688aaa193f332ed13e017e78ec07a7c80e45f6c5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_688aaa193f332ed13e017e78ec07a7c80e45f6c5.hip
deleted file mode 100644
index 0bd9a1cc08dee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_688aaa193f332ed13e017e78ec07a7c80e45f6c5.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6905ba47078abd7a5b6a51eb93b26095517e7f70.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6905ba47078abd7a5b6a51eb93b26095517e7f70.hip
deleted file mode 100644
index 204b0185fb129..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6905ba47078abd7a5b6a51eb93b26095517e7f70.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_69214eb450c3b249017480efb8d092b0edad6dc3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_69214eb450c3b249017480efb8d092b0edad6dc3.hip
deleted file mode 100644
index 7c8b6afe11854..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_69214eb450c3b249017480efb8d092b0edad6dc3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6979ef43adffdb62100270a62706fb811963925a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6979ef43adffdb62100270a62706fb811963925a.hip
deleted file mode 100644
index df832120ac3be..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6979ef43adffdb62100270a62706fb811963925a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_69cbe8eca7e3510f5caa7f13419cfbefbf031754.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_69cbe8eca7e3510f5caa7f13419cfbefbf031754.hip
deleted file mode 100644
index 0ab8867654657..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_69cbe8eca7e3510f5caa7f13419cfbefbf031754.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a3f42d5c9ccdd3807e488b00f02bc6ab5d8d99a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a3f42d5c9ccdd3807e488b00f02bc6ab5d8d99a.hip
deleted file mode 100644
index 510ebdb907393..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a3f42d5c9ccdd3807e488b00f02bc6ab5d8d99a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a4b6226b355bf35d4d07aaef1828091f03ad2ec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a4b6226b355bf35d4d07aaef1828091f03ad2ec.hip
deleted file mode 100644
index 9a3cd25c9722f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a4b6226b355bf35d4d07aaef1828091f03ad2ec.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a66604bb15f97a56847a7c968dbe32d247cbc13.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a66604bb15f97a56847a7c968dbe32d247cbc13.hip
deleted file mode 100644
index 29e2aae4eddbf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a66604bb15f97a56847a7c968dbe32d247cbc13.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a7b6781ffff9a42beebb4d73f0d15461ddd4479.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a7b6781ffff9a42beebb4d73f0d15461ddd4479.hip
deleted file mode 100644
index e8fe9491d28b2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a7b6781ffff9a42beebb4d73f0d15461ddd4479.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a7eb3d86aa385f9ecffbc5ba10489e56856f918.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a7eb3d86aa385f9ecffbc5ba10489e56856f918.hip
deleted file mode 100644
index 9859dbe06e07c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a7eb3d86aa385f9ecffbc5ba10489e56856f918.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a95543aeed81adfb6d847f78212585a36122ae3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a95543aeed81adfb6d847f78212585a36122ae3.hip
deleted file mode 100644
index 81ecc3b18d9e4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6a95543aeed81adfb6d847f78212585a36122ae3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6abeb7b50ae6a1fc62535b9a1dabbde6f177a9d0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6abeb7b50ae6a1fc62535b9a1dabbde6f177a9d0.hip
deleted file mode 100644
index 769a228a0473d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6abeb7b50ae6a1fc62535b9a1dabbde6f177a9d0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6af23d1460abfe875e71f7911697c42fef0f41c5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6af23d1460abfe875e71f7911697c42fef0f41c5.hip
deleted file mode 100644
index fa9915dd44a87..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6af23d1460abfe875e71f7911697c42fef0f41c5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6af4c15a119e805e4407b184625f57966f8833d9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6af4c15a119e805e4407b184625f57966f8833d9.hip
deleted file mode 100644
index 14534470a323d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6af4c15a119e805e4407b184625f57966f8833d9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6b0ef67ce0f178aa2863c4909f5bdd7f766c9b2f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6b0ef67ce0f178aa2863c4909f5bdd7f766c9b2f.hip
deleted file mode 100644
index 569684c725367..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6b0ef67ce0f178aa2863c4909f5bdd7f766c9b2f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6b638314efcc4f16aa4a6e58e6caf2fda1711519.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6b638314efcc4f16aa4a6e58e6caf2fda1711519.hip
deleted file mode 100644
index 750be296fb5a4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6b638314efcc4f16aa4a6e58e6caf2fda1711519.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6bad2ed9f91bc1efd89ea66cd5c775fa140cf931.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6bad2ed9f91bc1efd89ea66cd5c775fa140cf931.hip
deleted file mode 100644
index ad3022d6b9a38..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6bad2ed9f91bc1efd89ea66cd5c775fa140cf931.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6cfb7075345704340ff33dc0ef7c04ef127f26ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6cfb7075345704340ff33dc0ef7c04ef127f26ad.hip
deleted file mode 100644
index 05e91216be1af..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6cfb7075345704340ff33dc0ef7c04ef127f26ad.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d07bf9c05e41dcf2416e05dab4bdde17158db76.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d07bf9c05e41dcf2416e05dab4bdde17158db76.hip
deleted file mode 100644
index c4ab4808c8e08..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d07bf9c05e41dcf2416e05dab4bdde17158db76.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d17b92fab5bee7717bf9aff6a6bef7cee3816e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d17b92fab5bee7717bf9aff6a6bef7cee3816e7.hip
deleted file mode 100644
index b82ff8ab49835..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d17b92fab5bee7717bf9aff6a6bef7cee3816e7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d307974bdeeef95cca0d130ebb7aeb77fb1b6eb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d307974bdeeef95cca0d130ebb7aeb77fb1b6eb.hip
deleted file mode 100644
index 188d6d47f9e6b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d307974bdeeef95cca0d130ebb7aeb77fb1b6eb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d40d762ed576832b3a752453e9881b5fe6d2650.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d40d762ed576832b3a752453e9881b5fe6d2650.hip
deleted file mode 100644
index 2ddbefa0efe92..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d40d762ed576832b3a752453e9881b5fe6d2650.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d470f5c6fb81032fcd7974180297d4bb2a8427d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d470f5c6fb81032fcd7974180297d4bb2a8427d.hip
deleted file mode 100644
index 16d9abd5d89d1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d470f5c6fb81032fcd7974180297d4bb2a8427d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d5aad18f59e47a3fa3278c7ef1a6372830c33d5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d5aad18f59e47a3fa3278c7ef1a6372830c33d5.hip
deleted file mode 100644
index c56eecc5a3f2f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6d5aad18f59e47a3fa3278c7ef1a6372830c33d5.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6db86621d626722434f2ae9b7b8ab435a8dd8827.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6db86621d626722434f2ae9b7b8ab435a8dd8827.hip
deleted file mode 100644
index c5d84434223e6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6db86621d626722434f2ae9b7b8ab435a8dd8827.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6dd707cf48a17d31abef94215c5720419faa0a39.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6dd707cf48a17d31abef94215c5720419faa0a39.hip
deleted file mode 100644
index c609bd55973dc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6dd707cf48a17d31abef94215c5720419faa0a39.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e240106c771ebea461fc2a87b6da68e510aba70.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e240106c771ebea461fc2a87b6da68e510aba70.hip
deleted file mode 100644
index a288094482492..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e240106c771ebea461fc2a87b6da68e510aba70.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e6a4475ea795935f4cbf2dc0ac156a33d754587.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e6a4475ea795935f4cbf2dc0ac156a33d754587.hip
deleted file mode 100644
index 4ac93e1f60f2d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e6a4475ea795935f4cbf2dc0ac156a33d754587.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e7e1d245baabe2f6293e3d85318f9936b333500.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e7e1d245baabe2f6293e3d85318f9936b333500.hip
deleted file mode 100644
index 3948cba12a1fc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e7e1d245baabe2f6293e3d85318f9936b333500.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e8cda718e10824956f0ee39bbb0891eafa45a7b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e8cda718e10824956f0ee39bbb0891eafa45a7b.hip
deleted file mode 100644
index 4c427cbe2f812..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6e8cda718e10824956f0ee39bbb0891eafa45a7b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6eca9cd905ea8b0454cf9564643894682b08cb97.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6eca9cd905ea8b0454cf9564643894682b08cb97.hip
deleted file mode 100644
index b2e27eb1a9996..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6eca9cd905ea8b0454cf9564643894682b08cb97.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6eebd0c2fbfc85f938b10535855c388971129a28.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6eebd0c2fbfc85f938b10535855c388971129a28.hip
deleted file mode 100644
index 76d52c5061052..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6eebd0c2fbfc85f938b10535855c388971129a28.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ef5803b33d97db72eb8a8528aeb3fc956a938cc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ef5803b33d97db72eb8a8528aeb3fc956a938cc.hip
deleted file mode 100644
index 0d2092d2a3a27..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ef5803b33d97db72eb8a8528aeb3fc956a938cc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f31b3345893eec8ed1ddf1d8de2512b46ff6187.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f31b3345893eec8ed1ddf1d8de2512b46ff6187.hip
deleted file mode 100644
index dce7738d2c573..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f31b3345893eec8ed1ddf1d8de2512b46ff6187.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f3d098f8bb63133924aab70d26a6ed64018c13b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f3d098f8bb63133924aab70d26a6ed64018c13b.hip
deleted file mode 100644
index 07a4e63d703e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f3d098f8bb63133924aab70d26a6ed64018c13b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f8788c537cbf6833c58a6ca15c0a36de33c9fbd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f8788c537cbf6833c58a6ca15c0a36de33c9fbd.hip
deleted file mode 100644
index 3fe8cbefb4512..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f8788c537cbf6833c58a6ca15c0a36de33c9fbd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f88527a2cdb5adf51407f4661a254bb32d7de23.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f88527a2cdb5adf51407f4661a254bb32d7de23.hip
deleted file mode 100644
index 8fc6a93c12f12..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6f88527a2cdb5adf51407f4661a254bb32d7de23.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6fa6478cc27e52fd9511fbff38369c921155cfb9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6fa6478cc27e52fd9511fbff38369c921155cfb9.hip
deleted file mode 100644
index 26950a979caa3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6fa6478cc27e52fd9511fbff38369c921155cfb9.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ff4605d82507fc4bd6e96095eaee5173ea41973.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ff4605d82507fc4bd6e96095eaee5173ea41973.hip
deleted file mode 100644
index 9bbb2c9af8133..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ff4605d82507fc4bd6e96095eaee5173ea41973.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ff58a5186d69efd6062f3717bd315394ea6592b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ff58a5186d69efd6062f3717bd315394ea6592b.hip
deleted file mode 100644
index 0c7e48f516d20..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_6ff58a5186d69efd6062f3717bd315394ea6592b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_703246f1f53a988cf252eff88bdf814bd382d3ac.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_703246f1f53a988cf252eff88bdf814bd382d3ac.hip
deleted file mode 100644
index f499998516fde..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_703246f1f53a988cf252eff88bdf814bd382d3ac.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70586668a61ab88bc46b763df8f1c2ea52001ea0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70586668a61ab88bc46b763df8f1c2ea52001ea0.hip
deleted file mode 100644
index 1503096dd83cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70586668a61ab88bc46b763df8f1c2ea52001ea0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70c8e45f6ea7cf5dba9eeadd0b19481d9f5defb7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70c8e45f6ea7cf5dba9eeadd0b19481d9f5defb7.hip
deleted file mode 100644
index 96230751279bf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70c8e45f6ea7cf5dba9eeadd0b19481d9f5defb7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70cf755f1485c065222be4daab84283a9c3d0eb7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70cf755f1485c065222be4daab84283a9c3d0eb7.hip
deleted file mode 100644
index cb4aca8f61dc8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_70cf755f1485c065222be4daab84283a9c3d0eb7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_714c5369aa848021e020d874289e3ae4e0f74d77.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_714c5369aa848021e020d874289e3ae4e0f74d77.hip
deleted file mode 100644
index 400cdb6a7e5e9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_714c5369aa848021e020d874289e3ae4e0f74d77.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7177f939ac3dae8749cbf4232dcf04d2cf63b48f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7177f939ac3dae8749cbf4232dcf04d2cf63b48f.hip
deleted file mode 100644
index a8ac0905530e8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7177f939ac3dae8749cbf4232dcf04d2cf63b48f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71a2d046629a4b65c90d0e18d061c4984062f844.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71a2d046629a4b65c90d0e18d061c4984062f844.hip
deleted file mode 100644
index a96e18441bd64..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71a2d046629a4b65c90d0e18d061c4984062f844.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71b6100efe30d836dab557ea4ac54c4b9d35c6aa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71b6100efe30d836dab557ea4ac54c4b9d35c6aa.hip
deleted file mode 100644
index 65f5edef13d39..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71b6100efe30d836dab557ea4ac54c4b9d35c6aa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71dcbe9f481c92215f3b636bc0e86ce8f65e6472.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71dcbe9f481c92215f3b636bc0e86ce8f65e6472.hip
deleted file mode 100644
index 0a16cc4b5bbbf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71dcbe9f481c92215f3b636bc0e86ce8f65e6472.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71e3980331dc4bcec6ab6f4c345c7b5f71356979.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71e3980331dc4bcec6ab6f4c345c7b5f71356979.hip
deleted file mode 100644
index e8b3e80bf4049..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71e3980331dc4bcec6ab6f4c345c7b5f71356979.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71e5fb3544dafa9da03fd2de4bb9bd0718f6009f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71e5fb3544dafa9da03fd2de4bb9bd0718f6009f.hip
deleted file mode 100644
index d408c699f71a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_71e5fb3544dafa9da03fd2de4bb9bd0718f6009f.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7237ce5f3cf13ace3efc0b0227ae5a8c1fdfce1d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7237ce5f3cf13ace3efc0b0227ae5a8c1fdfce1d.hip
deleted file mode 100644
index 93bd7364e524c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7237ce5f3cf13ace3efc0b0227ae5a8c1fdfce1d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_724d1d4408196d611b2e0535bf8833652acbd6ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_724d1d4408196d611b2e0535bf8833652acbd6ef.hip
deleted file mode 100644
index a94b7aa89a80f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_724d1d4408196d611b2e0535bf8833652acbd6ef.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7264e378e1ea1d4dd97f6949d66f3492883b663e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7264e378e1ea1d4dd97f6949d66f3492883b663e.hip
deleted file mode 100644
index ae23481cdd20e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7264e378e1ea1d4dd97f6949d66f3492883b663e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_72abb25dba0c48b380b2dabeb6ab7efaa706d180.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_72abb25dba0c48b380b2dabeb6ab7efaa706d180.hip
deleted file mode 100644
index 974602b4240e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_72abb25dba0c48b380b2dabeb6ab7efaa706d180.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7309c38fc8a2d5ad6efd449107dc54a7509624fe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7309c38fc8a2d5ad6efd449107dc54a7509624fe.hip
deleted file mode 100644
index f931ad5c40d52..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7309c38fc8a2d5ad6efd449107dc54a7509624fe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7344f96bed2f56793b1c2583485aa161cdf30379.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7344f96bed2f56793b1c2583485aa161cdf30379.hip
deleted file mode 100644
index bbae0c6831c73..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7344f96bed2f56793b1c2583485aa161cdf30379.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7393267865f1c2b0aa1a09a586f54cec98eea4ae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7393267865f1c2b0aa1a09a586f54cec98eea4ae.hip
deleted file mode 100644
index a098096fb9959..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7393267865f1c2b0aa1a09a586f54cec98eea4ae.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_73d4901b8ef034590314048de7223a572d61ee0f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_73d4901b8ef034590314048de7223a572d61ee0f.hip
deleted file mode 100644
index c2f02318aab73..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_73d4901b8ef034590314048de7223a572d61ee0f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_73ec21ed6e040260c4f04ef68ef9307aa86985a7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_73ec21ed6e040260c4f04ef68ef9307aa86985a7.hip
deleted file mode 100644
index 201c917ac1d0e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_73ec21ed6e040260c4f04ef68ef9307aa86985a7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_741401abfbbbdf0dd1d62df8bc3e85371ead71d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_741401abfbbbdf0dd1d62df8bc3e85371ead71d6.hip
deleted file mode 100644
index 115428bbd3e6f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_741401abfbbbdf0dd1d62df8bc3e85371ead71d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_743176ecb1f0bc800c870861585edf56f88d7739.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_743176ecb1f0bc800c870861585edf56f88d7739.hip
deleted file mode 100644
index 22dcbcd3fa9c8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_743176ecb1f0bc800c870861585edf56f88d7739.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_744ec604c577a27e0aae5b39711a9e2eb82801b6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_744ec604c577a27e0aae5b39711a9e2eb82801b6.hip
deleted file mode 100644
index a65321800a916..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_744ec604c577a27e0aae5b39711a9e2eb82801b6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_745705ae121a1a331527cedfe4d31218a428a0df.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_745705ae121a1a331527cedfe4d31218a428a0df.hip
deleted file mode 100644
index 33f0dca78936a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_745705ae121a1a331527cedfe4d31218a428a0df.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_748a3d76e8ab73af9a5d2302d33e3b1d1b866dd1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_748a3d76e8ab73af9a5d2302d33e3b1d1b866dd1.hip
deleted file mode 100644
index e3c3b5c9ac18e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_748a3d76e8ab73af9a5d2302d33e3b1d1b866dd1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7497eca4d1a18306b406b367653622a8d64095bf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7497eca4d1a18306b406b367653622a8d64095bf.hip
deleted file mode 100644
index ee033b78fdec8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7497eca4d1a18306b406b367653622a8d64095bf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_74ba59d347ce8916a22b40e6f22a3c89e13db4d0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_74ba59d347ce8916a22b40e6f22a3c89e13db4d0.hip
deleted file mode 100644
index 4245769177b0f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_74ba59d347ce8916a22b40e6f22a3c89e13db4d0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_74d5f2aef029f2103bb419cc982cae99fd1a9253.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_74d5f2aef029f2103bb419cc982cae99fd1a9253.hip
deleted file mode 100644
index bfc786b112d62..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_74d5f2aef029f2103bb419cc982cae99fd1a9253.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7524904ac5a2040c7ea72aef5942212f291a21bf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7524904ac5a2040c7ea72aef5942212f291a21bf.hip
deleted file mode 100644
index df2a57dab01b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7524904ac5a2040c7ea72aef5942212f291a21bf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_758b211174da0f398b2a093e7389905b4f9c4060.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_758b211174da0f398b2a093e7389905b4f9c4060.hip
deleted file mode 100644
index 71fc61bebf4b3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_758b211174da0f398b2a093e7389905b4f9c4060.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7596c14b8fee751d03f42ca48ea4f66e87fc2e2f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7596c14b8fee751d03f42ca48ea4f66e87fc2e2f.hip
deleted file mode 100644
index 0276d8ab25513..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7596c14b8fee751d03f42ca48ea4f66e87fc2e2f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7597ce4d2e5264bdeda47487d5bdb55a014c6616.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7597ce4d2e5264bdeda47487d5bdb55a014c6616.hip
deleted file mode 100644
index 88817c29e52bd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7597ce4d2e5264bdeda47487d5bdb55a014c6616.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75a310a6eb86e3e8baac7a930c3ffbef372942b3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75a310a6eb86e3e8baac7a930c3ffbef372942b3.hip
deleted file mode 100644
index ce3f7983eac11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75a310a6eb86e3e8baac7a930c3ffbef372942b3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75c38912947881caa14b3fc7ab7bca317e296dc3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75c38912947881caa14b3fc7ab7bca317e296dc3.hip
deleted file mode 100644
index 83fc102c0a306..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75c38912947881caa14b3fc7ab7bca317e296dc3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75f2010bf6c478d2f0eba77e912697661306c1cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75f2010bf6c478d2f0eba77e912697661306c1cb.hip
deleted file mode 100644
index 7028223fc648e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75f2010bf6c478d2f0eba77e912697661306c1cb.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75f21e38ad01fade35b1db40adabd75eb602410c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75f21e38ad01fade35b1db40adabd75eb602410c.hip
deleted file mode 100644
index 4ef1fb0d6a349..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_75f21e38ad01fade35b1db40adabd75eb602410c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7601e6aea44b96e94fb019501be6b102c6e6a654.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7601e6aea44b96e94fb019501be6b102c6e6a654.hip
deleted file mode 100644
index a25ef56d4f8dc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7601e6aea44b96e94fb019501be6b102c6e6a654.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_761bde840c0c8149b24a8f6f264e963c4e9e8ceb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_761bde840c0c8149b24a8f6f264e963c4e9e8ceb.hip
deleted file mode 100644
index 5dfcc3c8d8e4c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_761bde840c0c8149b24a8f6f264e963c4e9e8ceb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_765940baaaa2ae6ade43ef4c94a220eaa63702b0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_765940baaaa2ae6ade43ef4c94a220eaa63702b0.hip
deleted file mode 100644
index 45dcf15438074..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_765940baaaa2ae6ade43ef4c94a220eaa63702b0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76674fc182dfa6329c73a354aa3adf458429444a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76674fc182dfa6329c73a354aa3adf458429444a.hip
deleted file mode 100644
index 4d6154fa41d53..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76674fc182dfa6329c73a354aa3adf458429444a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76704ca28a4877a1e84022e022614709adabb280.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76704ca28a4877a1e84022e022614709adabb280.hip
deleted file mode 100644
index d561dbf7a1cd7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76704ca28a4877a1e84022e022614709adabb280.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_768c80fd3ea17813df1bf19a158186834fd00780.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_768c80fd3ea17813df1bf19a158186834fd00780.hip
deleted file mode 100644
index 23abb80243975..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_768c80fd3ea17813df1bf19a158186834fd00780.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76be322fc072ca19baa82707e260c6eba936ae19.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76be322fc072ca19baa82707e260c6eba936ae19.hip
deleted file mode 100644
index e9a3df87b2207..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76be322fc072ca19baa82707e260c6eba936ae19.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76f884e9ca116ee47b446efe9fc770c178a858d5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76f884e9ca116ee47b446efe9fc770c178a858d5.hip
deleted file mode 100644
index e0ec95dc58f3e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_76f884e9ca116ee47b446efe9fc770c178a858d5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_770ad1eb1b30ad8f1e7c17df486093129b2d5630.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_770ad1eb1b30ad8f1e7c17df486093129b2d5630.hip
deleted file mode 100644
index 8643fee9bd74c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_770ad1eb1b30ad8f1e7c17df486093129b2d5630.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77200e875e0ef160b311c7de450c137772312d0d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77200e875e0ef160b311c7de450c137772312d0d.hip
deleted file mode 100644
index 73a829e1324cc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77200e875e0ef160b311c7de450c137772312d0d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_772016803aa3ca6ebe785557118365f9be7c4339.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_772016803aa3ca6ebe785557118365f9be7c4339.hip
deleted file mode 100644
index 6861aa9c8540a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_772016803aa3ca6ebe785557118365f9be7c4339.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7726be8909f631c04d4395fa4ffd03a736f447f1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7726be8909f631c04d4395fa4ffd03a736f447f1.hip
deleted file mode 100644
index 90de6047eabf8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7726be8909f631c04d4395fa4ffd03a736f447f1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7728d5bec7941c9b6d5632bee8d67ed92b9c03ec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7728d5bec7941c9b6d5632bee8d67ed92b9c03ec.hip
deleted file mode 100644
index 6eacd14761eaf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7728d5bec7941c9b6d5632bee8d67ed92b9c03ec.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7764814a0de7702f0b7b5ce9dede6440603f4853.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7764814a0de7702f0b7b5ce9dede6440603f4853.hip
deleted file mode 100644
index 240efadcbe2f4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7764814a0de7702f0b7b5ce9dede6440603f4853.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77a814291d8f01870274149b9d82fb75921d6e20.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77a814291d8f01870274149b9d82fb75921d6e20.hip
deleted file mode 100644
index 2c04d2cb994c9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77a814291d8f01870274149b9d82fb75921d6e20.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77d0223697ed41c4c2fd8830f8df6e5620db547f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77d0223697ed41c4c2fd8830f8df6e5620db547f.hip
deleted file mode 100644
index 7c3aca0d26a57..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_77d0223697ed41c4c2fd8830f8df6e5620db547f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7831ce329f2a0812ebb1dd103ea4ba8cb7ba531d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7831ce329f2a0812ebb1dd103ea4ba8cb7ba531d.hip
deleted file mode 100644
index 75adc7336ae53..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7831ce329f2a0812ebb1dd103ea4ba8cb7ba531d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7838849e57ee9cd292e588f587a8079b57becfc8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7838849e57ee9cd292e588f587a8079b57becfc8.hip
deleted file mode 100644
index 990ab970d7537..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7838849e57ee9cd292e588f587a8079b57becfc8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_783ec08544591a22f59dc12f169b7327b4185a1a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_783ec08544591a22f59dc12f169b7327b4185a1a.hip
deleted file mode 100644
index cb93457d2e4a2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_783ec08544591a22f59dc12f169b7327b4185a1a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_784c35fee4d372123631312f1051c43e1fa12378.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_784c35fee4d372123631312f1051c43e1fa12378.hip
deleted file mode 100644
index ac6ad19275d89..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_784c35fee4d372123631312f1051c43e1fa12378.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78663faeb0425f45e8a0da0f7b1a5ddbee5e07e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78663faeb0425f45e8a0da0f7b1a5ddbee5e07e7.hip
deleted file mode 100644
index 629c6dc42f66c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78663faeb0425f45e8a0da0f7b1a5ddbee5e07e7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7872c45ba170f2782c4b5b75cfc78ac79a4cf157.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7872c45ba170f2782c4b5b75cfc78ac79a4cf157.hip
deleted file mode 100644
index 0c76c45f5df76..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7872c45ba170f2782c4b5b75cfc78ac79a4cf157.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7878e2a4d3b96a552e03d1ffc33debfd50c9f7f1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7878e2a4d3b96a552e03d1ffc33debfd50c9f7f1.hip
deleted file mode 100644
index b63ad6a43b7da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7878e2a4d3b96a552e03d1ffc33debfd50c9f7f1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78e1edca5abe1bb3e7aa946eab6484b7bed806a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78e1edca5abe1bb3e7aa946eab6484b7bed806a3.hip
deleted file mode 100644
index e952938b0dbc8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78e1edca5abe1bb3e7aa946eab6484b7bed806a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78e945db4afa1330fe3978bc1bc9ae99828ae287.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78e945db4afa1330fe3978bc1bc9ae99828ae287.hip
deleted file mode 100644
index ad3c800147287..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78e945db4afa1330fe3978bc1bc9ae99828ae287.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78f7e2a2c08cd87702793f91b6935cbe4c22be55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78f7e2a2c08cd87702793f91b6935cbe4c22be55.hip
deleted file mode 100644
index 283fcca1f5b06..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_78f7e2a2c08cd87702793f91b6935cbe4c22be55.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_797750ac0b18b48f56ceb4640256e9bd3a36621a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_797750ac0b18b48f56ceb4640256e9bd3a36621a.hip
deleted file mode 100644
index ad4515d6a1505..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_797750ac0b18b48f56ceb4640256e9bd3a36621a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7993fc08ac5c6ce7a2eceb1227f4e3718dc4cf5f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7993fc08ac5c6ce7a2eceb1227f4e3718dc4cf5f.hip
deleted file mode 100644
index fd4a42f1a4030..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7993fc08ac5c6ce7a2eceb1227f4e3718dc4cf5f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79a7dce707954e765d97cb22e57d9bd6168860d9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79a7dce707954e765d97cb22e57d9bd6168860d9.hip
deleted file mode 100644
index d9d093e8665f7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79a7dce707954e765d97cb22e57d9bd6168860d9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79d0b8053ddf99a4d4447656d733c2da026b3a7c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79d0b8053ddf99a4d4447656d733c2da026b3a7c.hip
deleted file mode 100644
index 10636d6a616a9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79d0b8053ddf99a4d4447656d733c2da026b3a7c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79f182ae021e23869d7bebf2a9b4575bdc910ed0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79f182ae021e23869d7bebf2a9b4575bdc910ed0.hip
deleted file mode 100644
index 457b54d3dace7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_79f182ae021e23869d7bebf2a9b4575bdc910ed0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a0ab620e6d62259a559e329460e46e6e3f7c3f9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a0ab620e6d62259a559e329460e46e6e3f7c3f9.hip
deleted file mode 100644
index 74bb83256b39a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a0ab620e6d62259a559e329460e46e6e3f7c3f9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a13d62a715fd717f0d4101f787349cb49cbe70f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a13d62a715fd717f0d4101f787349cb49cbe70f.hip
deleted file mode 100644
index 53418d562863e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a13d62a715fd717f0d4101f787349cb49cbe70f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a242e5953f44316b6a4f6587ec26283ed6cbcae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a242e5953f44316b6a4f6587ec26283ed6cbcae.hip
deleted file mode 100644
index d75e6e191f6ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a242e5953f44316b6a4f6587ec26283ed6cbcae.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a2e032f6500fbc5468183415b6dd1d3e43f0bee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a2e032f6500fbc5468183415b6dd1d3e43f0bee.hip
deleted file mode 100644
index dff447dd2ce24..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a2e032f6500fbc5468183415b6dd1d3e43f0bee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a890b126da2d8cfbf84f048b779cac2dd56b509.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a890b126da2d8cfbf84f048b779cac2dd56b509.hip
deleted file mode 100644
index 88d05d2d89649..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a890b126da2d8cfbf84f048b779cac2dd56b509.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a902ed4ae3cc6558c73b730ff3949778007a230.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a902ed4ae3cc6558c73b730ff3949778007a230.hip
deleted file mode 100644
index cae30bbc80276..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7a902ed4ae3cc6558c73b730ff3949778007a230.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7aa14aa94d625b33df1adfa30ef4d91769592608.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7aa14aa94d625b33df1adfa30ef4d91769592608.hip
deleted file mode 100644
index 3138c9f5a7e70..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7aa14aa94d625b33df1adfa30ef4d91769592608.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ab03a62e064864e1e9c1cd506c1b2e1786a777c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ab03a62e064864e1e9c1cd506c1b2e1786a777c.hip
deleted file mode 100644
index 606cca5dd072c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ab03a62e064864e1e9c1cd506c1b2e1786a777c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7adf69b51f0a8cc9ae7e250e60df38758230fe4f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7adf69b51f0a8cc9ae7e250e60df38758230fe4f.hip
deleted file mode 100644
index 26a042289b479..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7adf69b51f0a8cc9ae7e250e60df38758230fe4f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7afd1a756247b15b078d15a39e350a07c22982da.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7afd1a756247b15b078d15a39e350a07c22982da.hip
deleted file mode 100644
index 3ad5895b74399..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7afd1a756247b15b078d15a39e350a07c22982da.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b2d3680c3578c7292349b58843aef7a82e0087d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b2d3680c3578c7292349b58843aef7a82e0087d.hip
deleted file mode 100644
index 179b17c14a7ba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b2d3680c3578c7292349b58843aef7a82e0087d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b5680f97836be4a369802e8115617a83875703e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b5680f97836be4a369802e8115617a83875703e.hip
deleted file mode 100644
index 55123992c1cfe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b5680f97836be4a369802e8115617a83875703e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b67045d438a7e4b8f3a313a5df5a85f351c1be5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b67045d438a7e4b8f3a313a5df5a85f351c1be5.hip
deleted file mode 100644
index a76f03ddda752..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b67045d438a7e4b8f3a313a5df5a85f351c1be5.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b7fa76609243a8709f349ffc0d9d88157f28dc9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b7fa76609243a8709f349ffc0d9d88157f28dc9.hip
deleted file mode 100644
index 60b1d61672e31..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b7fa76609243a8709f349ffc0d9d88157f28dc9.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b9a3bf1a9b37e0bd9bae6249609e5994dc0dba1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b9a3bf1a9b37e0bd9bae6249609e5994dc0dba1.hip
deleted file mode 100644
index 11548e314ba22..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7b9a3bf1a9b37e0bd9bae6249609e5994dc0dba1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7bb7b63e8a4c1df4eac4d978e166867195bd6e53.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7bb7b63e8a4c1df4eac4d978e166867195bd6e53.hip
deleted file mode 100644
index 3f0066d8e6566..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7bb7b63e8a4c1df4eac4d978e166867195bd6e53.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c19fc90e5a9c422dbf529d2def286f47dea0f50.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c19fc90e5a9c422dbf529d2def286f47dea0f50.hip
deleted file mode 100644
index dbac69a6daa25..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c19fc90e5a9c422dbf529d2def286f47dea0f50.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c23dde1a386436e9864c8fa5f1706c0d2fbfd0d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c23dde1a386436e9864c8fa5f1706c0d2fbfd0d.hip
deleted file mode 100644
index a5529a6cccaea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c23dde1a386436e9864c8fa5f1706c0d2fbfd0d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c3d8ef4da515960bf40eb1feb04d21950ad5ae5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c3d8ef4da515960bf40eb1feb04d21950ad5ae5.hip
deleted file mode 100644
index d3433f804b6b7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c3d8ef4da515960bf40eb1feb04d21950ad5ae5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c4710e8f4e27fae4ae079f1667c3a1879cb6da8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c4710e8f4e27fae4ae079f1667c3a1879cb6da8.hip
deleted file mode 100644
index 1114a63ffb4f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7c4710e8f4e27fae4ae079f1667c3a1879cb6da8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7cbe4562c51d6829ec5942e11035c452fe318b3a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7cbe4562c51d6829ec5942e11035c452fe318b3a.hip
deleted file mode 100644
index 75ff01bb255a6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7cbe4562c51d6829ec5942e11035c452fe318b3a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7cdc419d4248dfdeeab1f0980aec35fa134e52e0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7cdc419d4248dfdeeab1f0980aec35fa134e52e0.hip
deleted file mode 100644
index b727c425b11c2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7cdc419d4248dfdeeab1f0980aec35fa134e52e0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d08373ace7087bdaca4ce8b0bc329f553f88d77.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d08373ace7087bdaca4ce8b0bc329f553f88d77.hip
deleted file mode 100644
index eb5e4cb696608..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d08373ace7087bdaca4ce8b0bc329f553f88d77.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d0f767c17385eb7d756cbe8ed444d7cef72dea5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d0f767c17385eb7d756cbe8ed444d7cef72dea5.hip
deleted file mode 100644
index 31b5a9b567d20..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d0f767c17385eb7d756cbe8ed444d7cef72dea5.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d12e9cb599d24631c082e3cf65d2c58b6d4d44f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d12e9cb599d24631c082e3cf65d2c58b6d4d44f.hip
deleted file mode 100644
index f781a6ffc6104..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d12e9cb599d24631c082e3cf65d2c58b6d4d44f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d2f87c021e0b6a27b2d7e30351fd50f06414b5f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d2f87c021e0b6a27b2d7e30351fd50f06414b5f.hip
deleted file mode 100644
index 28b073a0087f3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d2f87c021e0b6a27b2d7e30351fd50f06414b5f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d5667b27f15a06d4040354fba3601d48bb9c045.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d5667b27f15a06d4040354fba3601d48bb9c045.hip
deleted file mode 100644
index a82404cc00930..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7d5667b27f15a06d4040354fba3601d48bb9c045.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dac5d4cf103d658e129673549549f1276f134e0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dac5d4cf103d658e129673549549f1276f134e0.hip
deleted file mode 100644
index 2f58ff3a4fe06..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dac5d4cf103d658e129673549549f1276f134e0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dd260849b86c46b685955cab54ba07d49b47954.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dd260849b86c46b685955cab54ba07d49b47954.hip
deleted file mode 100644
index 5cf98e1f5e42e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dd260849b86c46b685955cab54ba07d49b47954.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ddd621da88c57798db1e689b93b692b6519ff96.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ddd621da88c57798db1e689b93b692b6519ff96.hip
deleted file mode 100644
index 3a15564be0986..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ddd621da88c57798db1e689b93b692b6519ff96.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dfe21ee27f8a0ca0407ef0dea73cd73ae6940db.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dfe21ee27f8a0ca0407ef0dea73cd73ae6940db.hip
deleted file mode 100644
index 1ef8422004e9a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7dfe21ee27f8a0ca0407ef0dea73cd73ae6940db.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e1bdde812c332c9fc58613698568a04771b9fa8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e1bdde812c332c9fc58613698568a04771b9fa8.hip
deleted file mode 100644
index f0199b16708b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e1bdde812c332c9fc58613698568a04771b9fa8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e332a6aeecfb12dcf70c69157fd3137343fb9f6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e332a6aeecfb12dcf70c69157fd3137343fb9f6.hip
deleted file mode 100644
index 4397946dadae0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e332a6aeecfb12dcf70c69157fd3137343fb9f6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e6129eead18d13a4a6cb9550384fddabc7a2a16.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e6129eead18d13a4a6cb9550384fddabc7a2a16.hip
deleted file mode 100644
index c009b505f4f0a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e6129eead18d13a4a6cb9550384fddabc7a2a16.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e89f79217037e361bb0909d06534e40f5026b4f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e89f79217037e361bb0909d06534e40f5026b4f.hip
deleted file mode 100644
index 4739379775d96..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e89f79217037e361bb0909d06534e40f5026b4f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e9519dd0d0f940fd5efd61bd32df7528ba7e3fc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e9519dd0d0f940fd5efd61bd32df7528ba7e3fc.hip
deleted file mode 100644
index ea24e86a28f30..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e9519dd0d0f940fd5efd61bd32df7528ba7e3fc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e9c7feb747241c9c7de2adf3a19933a1c4c0995.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e9c7feb747241c9c7de2adf3a19933a1c4c0995.hip
deleted file mode 100644
index 88cabf351126e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7e9c7feb747241c9c7de2adf3a19933a1c4c0995.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ea9c37d92e344f3cc58cd4d1d00f19167e3623e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ea9c37d92e344f3cc58cd4d1d00f19167e3623e.hip
deleted file mode 100644
index 936d1899f4d60..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ea9c37d92e344f3cc58cd4d1d00f19167e3623e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ec038393ec329a894aee9bbac078a40f57a4684.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ec038393ec329a894aee9bbac078a40f57a4684.hip
deleted file mode 100644
index a2324dfee77c0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ec038393ec329a894aee9bbac078a40f57a4684.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ec04763d635c5bc3e810737b5d948c59f117d5a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ec04763d635c5bc3e810737b5d948c59f117d5a.hip
deleted file mode 100644
index 259c60ca13c09..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ec04763d635c5bc3e810737b5d948c59f117d5a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ee953cb24e28bcdc8f05783894b23cbf83bdf35.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ee953cb24e28bcdc8f05783894b23cbf83bdf35.hip
deleted file mode 100644
index 117ed8c59c353..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ee953cb24e28bcdc8f05783894b23cbf83bdf35.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f6ccdb3c2d595fffd05bc5e6417b157276547fb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f6ccdb3c2d595fffd05bc5e6417b157276547fb.hip
deleted file mode 100644
index d6e33f0a43ed4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f6ccdb3c2d595fffd05bc5e6417b157276547fb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f80d44e82e601dc48d4c8b4e710ef7265894b6c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f80d44e82e601dc48d4c8b4e710ef7265894b6c.hip
deleted file mode 100644
index 04eab9f3a2c94..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f80d44e82e601dc48d4c8b4e710ef7265894b6c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f9403cb91d6aabebf081afae94a8ba397d8d24f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f9403cb91d6aabebf081afae94a8ba397d8d24f.hip
deleted file mode 100644
index e1c494516ffc7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f9403cb91d6aabebf081afae94a8ba397d8d24f.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f9bb3486fee7b7c9e24300b8a4e4ce88a11bfc0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f9bb3486fee7b7c9e24300b8a4e4ce88a11bfc0.hip
deleted file mode 100644
index 0139958c74591..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7f9bb3486fee7b7c9e24300b8a4e4ce88a11bfc0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7fa76fc1b066a15b08dc6c24a7cf33a58b4cb6cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7fa76fc1b066a15b08dc6c24a7cf33a58b4cb6cb.hip
deleted file mode 100644
index 8a13f4c6e39d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7fa76fc1b066a15b08dc6c24a7cf33a58b4cb6cb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7fe409f4421193fb48a54aa5f26bd6229d23204c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7fe409f4421193fb48a54aa5f26bd6229d23204c.hip
deleted file mode 100644
index 58d6c86822d55..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7fe409f4421193fb48a54aa5f26bd6229d23204c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ff65c7abd9b0d8a2df9302d6dc167637b3a72f0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ff65c7abd9b0d8a2df9302d6dc167637b3a72f0.hip
deleted file mode 100644
index a408de6792905..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_7ff65c7abd9b0d8a2df9302d6dc167637b3a72f0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8004763f674dfb3f14b66dfdeb2a046e413ce2cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8004763f674dfb3f14b66dfdeb2a046e413ce2cb.hip
deleted file mode 100644
index 5e9b6e48392df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8004763f674dfb3f14b66dfdeb2a046e413ce2cb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8007bf7ae1b71bf8ac4a793aa519ad333aa7a7ba.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8007bf7ae1b71bf8ac4a793aa519ad333aa7a7ba.hip
deleted file mode 100644
index 3e015f8f7029d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8007bf7ae1b71bf8ac4a793aa519ad333aa7a7ba.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8021fa266c77e6b5bd1af2a9c22c686e5a6eac78.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8021fa266c77e6b5bd1af2a9c22c686e5a6eac78.hip
deleted file mode 100644
index ac28a33284680..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8021fa266c77e6b5bd1af2a9c22c686e5a6eac78.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_802b21f9588d72c3c3e3b9a3b269f19c484d5aa4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_802b21f9588d72c3c3e3b9a3b269f19c484d5aa4.hip
deleted file mode 100644
index 4b60fc9210e03..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_802b21f9588d72c3c3e3b9a3b269f19c484d5aa4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8046f566fa7188c92568b277354e8b06ad382544.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8046f566fa7188c92568b277354e8b06ad382544.hip
deleted file mode 100644
index c215fe79952fb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8046f566fa7188c92568b277354e8b06ad382544.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_806f9ab9baf631df1d3a8d801e4cf93a102526cf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_806f9ab9baf631df1d3a8d801e4cf93a102526cf.hip
deleted file mode 100644
index 15f6db450b18f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_806f9ab9baf631df1d3a8d801e4cf93a102526cf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_807545400aa6e70ff49a5f38ed6a218a180bd87f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_807545400aa6e70ff49a5f38ed6a218a180bd87f.hip
deleted file mode 100644
index 1a617692a47f4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_807545400aa6e70ff49a5f38ed6a218a180bd87f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80987e2d765efc320eaee813607c94c80ee35aa4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80987e2d765efc320eaee813607c94c80ee35aa4.hip
deleted file mode 100644
index 3f19fb11a2f4d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80987e2d765efc320eaee813607c94c80ee35aa4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80a72d70d80b66c19e85daa00497308381050048.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80a72d70d80b66c19e85daa00497308381050048.hip
deleted file mode 100644
index 61ad99831392b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80a72d70d80b66c19e85daa00497308381050048.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80bfb0e6032892cc58cef4dd403f305a5b76851b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80bfb0e6032892cc58cef4dd403f305a5b76851b.hip
deleted file mode 100644
index 086ab0ba7a54c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80bfb0e6032892cc58cef4dd403f305a5b76851b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80cf0997573f4bcfbaaf75e40f519580a7495a17.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80cf0997573f4bcfbaaf75e40f519580a7495a17.hip
deleted file mode 100644
index c7eff142a038d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80cf0997573f4bcfbaaf75e40f519580a7495a17.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80efc341089a50ed5669b3c86f6ddd9b124d1442.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80efc341089a50ed5669b3c86f6ddd9b124d1442.hip
deleted file mode 100644
index 4cbab736b6b51..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80efc341089a50ed5669b3c86f6ddd9b124d1442.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80f51f0e178c33e6196df1d2e47bd38bf5391cc8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80f51f0e178c33e6196df1d2e47bd38bf5391cc8.hip
deleted file mode 100644
index 66cd7a461f77a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80f51f0e178c33e6196df1d2e47bd38bf5391cc8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80fb694fce7b4c3c459fca43c89c6002fbfdaef5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80fb694fce7b4c3c459fca43c89c6002fbfdaef5.hip
deleted file mode 100644
index 2658df162189f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_80fb694fce7b4c3c459fca43c89c6002fbfdaef5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_810dd4e870ceda3ba9b5f0084a4b025b2e609d57.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_810dd4e870ceda3ba9b5f0084a4b025b2e609d57.hip
deleted file mode 100644
index 4ab0c214f0b09..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_810dd4e870ceda3ba9b5f0084a4b025b2e609d57.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_811db756577b61cde9fe8279d956980db9ee21a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_811db756577b61cde9fe8279d956980db9ee21a4.hip
deleted file mode 100644
index de717ef0b7a77..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_811db756577b61cde9fe8279d956980db9ee21a4.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_813e60e8405aca3f7fbed19452ae37574ada9a77.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_813e60e8405aca3f7fbed19452ae37574ada9a77.hip
deleted file mode 100644
index 33aae40fecb28..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_813e60e8405aca3f7fbed19452ae37574ada9a77.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_815918206483d2ae04a45aa67d69dfb986587214.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_815918206483d2ae04a45aa67d69dfb986587214.hip
deleted file mode 100644
index b6e2d5b3410e3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_815918206483d2ae04a45aa67d69dfb986587214.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_816c48e129a0235cb3a19124ddb28cce286fb368.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_816c48e129a0235cb3a19124ddb28cce286fb368.hip
deleted file mode 100644
index fb3b8e71300bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_816c48e129a0235cb3a19124ddb28cce286fb368.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81acf1d17650712b71a499bb66909bfcfcb6aecb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81acf1d17650712b71a499bb66909bfcfcb6aecb.hip
deleted file mode 100644
index 219fd7dd1c9c1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81acf1d17650712b71a499bb66909bfcfcb6aecb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81bb8f13b6f20a72c9ce6d0b53f81eddbf05f1c6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81bb8f13b6f20a72c9ce6d0b53f81eddbf05f1c6.hip
deleted file mode 100644
index 1b86e63fc6e0c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81bb8f13b6f20a72c9ce6d0b53f81eddbf05f1c6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81dd3ea61bb61de02667b14f5a94198f48c7307b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81dd3ea61bb61de02667b14f5a94198f48c7307b.hip
deleted file mode 100644
index dfe9501b99a70..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81dd3ea61bb61de02667b14f5a94198f48c7307b.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81f6c575c3fa2ccc7e65022f1ba65c8cfc16541e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81f6c575c3fa2ccc7e65022f1ba65c8cfc16541e.hip
deleted file mode 100644
index fab9835e1c0ca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_81f6c575c3fa2ccc7e65022f1ba65c8cfc16541e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82048cf91270631f98ac37dc488a1fb2e00ce004.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82048cf91270631f98ac37dc488a1fb2e00ce004.hip
deleted file mode 100644
index 9d110471afdb9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82048cf91270631f98ac37dc488a1fb2e00ce004.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8250f27341241086515d833aa53ae873d4ece3fa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8250f27341241086515d833aa53ae873d4ece3fa.hip
deleted file mode 100644
index 7b2f9217aa6b8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8250f27341241086515d833aa53ae873d4ece3fa.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8278845045d68027dcf3bf867ecde2fb12ec51d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8278845045d68027dcf3bf867ecde2fb12ec51d3.hip
deleted file mode 100644
index bf38e35bfca18..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8278845045d68027dcf3bf867ecde2fb12ec51d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82ad0c0580516485ea432d98f53e73f6dfec548c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82ad0c0580516485ea432d98f53e73f6dfec548c.hip
deleted file mode 100644
index 2b4150ecc8ee1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82ad0c0580516485ea432d98f53e73f6dfec548c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82c932e6eaaf44861c794539d9caf8b50192fc44.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82c932e6eaaf44861c794539d9caf8b50192fc44.hip
deleted file mode 100644
index 8b2eed5f8005c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82c932e6eaaf44861c794539d9caf8b50192fc44.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82d7f61e6313930f063758b61102e7a43b118beb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82d7f61e6313930f063758b61102e7a43b118beb.hip
deleted file mode 100644
index 8f7f3de158bac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82d7f61e6313930f063758b61102e7a43b118beb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82f0f3d71108dcc49234a258f0f3b21ea2123cc0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82f0f3d71108dcc49234a258f0f3b21ea2123cc0.hip
deleted file mode 100644
index 93daa709c4499..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82f0f3d71108dcc49234a258f0f3b21ea2123cc0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82f1d7e1a93bf2fa80c409e6827ea88af56c44f0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82f1d7e1a93bf2fa80c409e6827ea88af56c44f0.hip
deleted file mode 100644
index 54f9ec27a3fe7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_82f1d7e1a93bf2fa80c409e6827ea88af56c44f0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8301bfc0394936a68fa0098580f06e77c88ebed9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8301bfc0394936a68fa0098580f06e77c88ebed9.hip
deleted file mode 100644
index 4cf9f48de04e3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8301bfc0394936a68fa0098580f06e77c88ebed9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83080406598df6bd3102db70a554e496e29db96a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83080406598df6bd3102db70a554e496e29db96a.hip
deleted file mode 100644
index 3144707a1e30b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83080406598df6bd3102db70a554e496e29db96a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_830e3532f27b391585d5de90f3bdf97992b67651.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_830e3532f27b391585d5de90f3bdf97992b67651.hip
deleted file mode 100644
index 6c320947ef899..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_830e3532f27b391585d5de90f3bdf97992b67651.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8352031044ef2e4a22e27ad04ab5d2c02121faee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8352031044ef2e4a22e27ad04ab5d2c02121faee.hip
deleted file mode 100644
index 88b13709d5ec2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8352031044ef2e4a22e27ad04ab5d2c02121faee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_835a906031a258c6362313eec783678bd8125c91.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_835a906031a258c6362313eec783678bd8125c91.hip
deleted file mode 100644
index e0a251e22ddd2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_835a906031a258c6362313eec783678bd8125c91.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_836a308c2d2afd6e0dfbfda61984b631c4ccffc6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_836a308c2d2afd6e0dfbfda61984b631c4ccffc6.hip
deleted file mode 100644
index 2c4cdbac3707c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_836a308c2d2afd6e0dfbfda61984b631c4ccffc6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83d580a612af85533c87aecdd7b0345c71b75980.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83d580a612af85533c87aecdd7b0345c71b75980.hip
deleted file mode 100644
index 306db4dc5008b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83d580a612af85533c87aecdd7b0345c71b75980.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83d920a76114c63156740ba5dd6f3846c4b21c28.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83d920a76114c63156740ba5dd6f3846c4b21c28.hip
deleted file mode 100644
index 204b6dee77714..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83d920a76114c63156740ba5dd6f3846c4b21c28.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83ddca2c6ecbba4314c434e7471ffb8fa642f936.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83ddca2c6ecbba4314c434e7471ffb8fa642f936.hip
deleted file mode 100644
index ac353df75b245..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83ddca2c6ecbba4314c434e7471ffb8fa642f936.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83f6a1837a65df12b7c55d25ca28cc939c2a6328.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83f6a1837a65df12b7c55d25ca28cc939c2a6328.hip
deleted file mode 100644
index e0f7b733dadce..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_83f6a1837a65df12b7c55d25ca28cc939c2a6328.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_843e7888cba5f463d19fcb71aaaab25dc3d2c09d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_843e7888cba5f463d19fcb71aaaab25dc3d2c09d.hip
deleted file mode 100644
index 10012bf1b8dcc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_843e7888cba5f463d19fcb71aaaab25dc3d2c09d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8441910c34830ad2459fb85c2c14af02da718fdc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8441910c34830ad2459fb85c2c14af02da718fdc.hip
deleted file mode 100644
index 09f65ddffee95..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8441910c34830ad2459fb85c2c14af02da718fdc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8457ea5726149efb8778e6d90798b8e48288fc9a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8457ea5726149efb8778e6d90798b8e48288fc9a.hip
deleted file mode 100644
index 1168a111bfc3d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8457ea5726149efb8778e6d90798b8e48288fc9a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_847feaf237911478173377a501ee19ee325b012b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_847feaf237911478173377a501ee19ee325b012b.hip
deleted file mode 100644
index fb8c2ff465eab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_847feaf237911478173377a501ee19ee325b012b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84cca7528c7d1bf49ba79625733ff0ae7522c096.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84cca7528c7d1bf49ba79625733ff0ae7522c096.hip
deleted file mode 100644
index 09957d445dcf9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84cca7528c7d1bf49ba79625733ff0ae7522c096.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84dc4af43de08130a04bfa06df9799b6e9e96900.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84dc4af43de08130a04bfa06df9799b6e9e96900.hip
deleted file mode 100644
index 5bcfa28a7c216..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84dc4af43de08130a04bfa06df9799b6e9e96900.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84e8ae99e184013739019c93d07caddce532382b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84e8ae99e184013739019c93d07caddce532382b.hip
deleted file mode 100644
index 14b46ad67ae2a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84e8ae99e184013739019c93d07caddce532382b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84fc5e94f89d6a9287cf64662a372784511468dd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84fc5e94f89d6a9287cf64662a372784511468dd.hip
deleted file mode 100644
index a6bef398368ab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_84fc5e94f89d6a9287cf64662a372784511468dd.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8513d96a66a4d9fb8dfc84afba7e1d8c200248a6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8513d96a66a4d9fb8dfc84afba7e1d8c200248a6.hip
deleted file mode 100644
index d99e00fd35a8c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8513d96a66a4d9fb8dfc84afba7e1d8c200248a6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85156f2c556c6ef6180608c361b7b35ede71ffea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85156f2c556c6ef6180608c361b7b35ede71ffea.hip
deleted file mode 100644
index 785187122a1f7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85156f2c556c6ef6180608c361b7b35ede71ffea.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_854c8003a508ed3f8cbe6967c4ae2635a491c721.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_854c8003a508ed3f8cbe6967c4ae2635a491c721.hip
deleted file mode 100644
index b44495ac38bb1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_854c8003a508ed3f8cbe6967c4ae2635a491c721.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85908fe6dc9c629c82d6953081b10021e64583b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85908fe6dc9c629c82d6953081b10021e64583b1.hip
deleted file mode 100644
index 7ecdf21a02d79..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85908fe6dc9c629c82d6953081b10021e64583b1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85960fe542635079de5eca3c7785890cd4740005.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85960fe542635079de5eca3c7785890cd4740005.hip
deleted file mode 100644
index 849aa28fc080b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85960fe542635079de5eca3c7785890cd4740005.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85fdde4b25e2fc8cbdd46c2850c19eac8d9af8f6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85fdde4b25e2fc8cbdd46c2850c19eac8d9af8f6.hip
deleted file mode 100644
index 00be09d0ffcc6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_85fdde4b25e2fc8cbdd46c2850c19eac8d9af8f6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86309c036d96367939ccc3e8922595ac35a3e179.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86309c036d96367939ccc3e8922595ac35a3e179.hip
deleted file mode 100644
index 8057a5c80445b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86309c036d96367939ccc3e8922595ac35a3e179.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86513d6e065a44bcb0c789eed1e7e5456e800ab6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86513d6e065a44bcb0c789eed1e7e5456e800ab6.hip
deleted file mode 100644
index cb886078cd819..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86513d6e065a44bcb0c789eed1e7e5456e800ab6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_865eb90b1a2d64acc0f6fbe1d807c501fd4be3cd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_865eb90b1a2d64acc0f6fbe1d807c501fd4be3cd.hip
deleted file mode 100644
index f970343ef9548..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_865eb90b1a2d64acc0f6fbe1d807c501fd4be3cd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8689126a7eb09d81baaf8f99dbff8932fbeab3cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8689126a7eb09d81baaf8f99dbff8932fbeab3cb.hip
deleted file mode 100644
index 8a835605072c5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8689126a7eb09d81baaf8f99dbff8932fbeab3cb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86d73393d0d8b769f30222f7817563a955c36dfc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86d73393d0d8b769f30222f7817563a955c36dfc.hip
deleted file mode 100644
index 1ae87b99d6fc0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86d73393d0d8b769f30222f7817563a955c36dfc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86fa51b8c7a2f3fac5cf4cd2951ed2ede5c35450.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86fa51b8c7a2f3fac5cf4cd2951ed2ede5c35450.hip
deleted file mode 100644
index 5211cf9c9bea3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_86fa51b8c7a2f3fac5cf4cd2951ed2ede5c35450.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_875b08ca602fe48840c72cd61798acb98540fcd6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_875b08ca602fe48840c72cd61798acb98540fcd6.hip
deleted file mode 100644
index f484bc9e71bf5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_875b08ca602fe48840c72cd61798acb98540fcd6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_876a418fbe6183d0392b7a7d9986d067e323e2b9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_876a418fbe6183d0392b7a7d9986d067e323e2b9.hip
deleted file mode 100644
index f33f73433e6f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_876a418fbe6183d0392b7a7d9986d067e323e2b9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_877e33463b3bf1853c6d2d2009af8d27bf88abbe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_877e33463b3bf1853c6d2d2009af8d27bf88abbe.hip
deleted file mode 100644
index 20166c479009b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_877e33463b3bf1853c6d2d2009af8d27bf88abbe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8793dc3217e154b65ebba065aa10ab4dc2374ae8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8793dc3217e154b65ebba065aa10ab4dc2374ae8.hip
deleted file mode 100644
index 745fa19f66d9f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8793dc3217e154b65ebba065aa10ab4dc2374ae8.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::fp16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_87e3a06266deda093bdf28af82d8666066157fc6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_87e3a06266deda093bdf28af82d8666066157fc6.hip
deleted file mode 100644
index 43ea7e1a8a245..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_87e3a06266deda093bdf28af82d8666066157fc6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8840e8899b4e632714632450bcef001c6070f955.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8840e8899b4e632714632450bcef001c6070f955.hip
deleted file mode 100644
index e83e3c95bb813..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8840e8899b4e632714632450bcef001c6070f955.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ac7f6cbdfca2e397bcb86af4216e87166601c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ac7f6cbdfca2e397bcb86af4216e87166601c7.hip
deleted file mode 100644
index fa4108d06f38e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ac7f6cbdfca2e397bcb86af4216e87166601c7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88c04463f9c5ce565a9daa8c22e16de80fadd707.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88c04463f9c5ce565a9daa8c22e16de80fadd707.hip
deleted file mode 100644
index 282c11cb4bec4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88c04463f9c5ce565a9daa8c22e16de80fadd707.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88d52c5f70abb525b9c8aa8fc1cb3997c33ed67c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88d52c5f70abb525b9c8aa8fc1cb3997c33ed67c.hip
deleted file mode 100644
index ef926d8790999..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88d52c5f70abb525b9c8aa8fc1cb3997c33ed67c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ea5b5346c87cc4fc1e841c518080df4ab811a2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ea5b5346c87cc4fc1e841c518080df4ab811a2.hip
deleted file mode 100644
index 7717baf29b11f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ea5b5346c87cc4fc1e841c518080df4ab811a2.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ed7f650c958a644c8031aeb88688b1e42458e5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ed7f650c958a644c8031aeb88688b1e42458e5.hip
deleted file mode 100644
index 171f415499bbe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_88ed7f650c958a644c8031aeb88688b1e42458e5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_890aa875ac13957f00b30210477924697abf0c9e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_890aa875ac13957f00b30210477924697abf0c9e.hip
deleted file mode 100644
index d4342466be28c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_890aa875ac13957f00b30210477924697abf0c9e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_89617bdea526d12d6a33ed42b9b0018c0b173722.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_89617bdea526d12d6a33ed42b9b0018c0b173722.hip
deleted file mode 100644
index e1f24ddd6cdf4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_89617bdea526d12d6a33ed42b9b0018c0b173722.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_89a3327da9a3411ff1cddc67eb647083cd947a92.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_89a3327da9a3411ff1cddc67eb647083cd947a92.hip
deleted file mode 100644
index 7d4b97f05e36c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_89a3327da9a3411ff1cddc67eb647083cd947a92.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a1fd28acfe85b3adac859c4bbffa4d28fe634fe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a1fd28acfe85b3adac859c4bbffa4d28fe634fe.hip
deleted file mode 100644
index 8a57e89bea40f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a1fd28acfe85b3adac859c4bbffa4d28fe634fe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a58d4bca33c4c0e79141a56688049237d170d1b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a58d4bca33c4c0e79141a56688049237d170d1b.hip
deleted file mode 100644
index 7f79025e61a5f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a58d4bca33c4c0e79141a56688049237d170d1b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a824621a50cdc3cbadc4b1f9ef18e1325385082.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a824621a50cdc3cbadc4b1f9ef18e1325385082.hip
deleted file mode 100644
index 030234915a54a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a824621a50cdc3cbadc4b1f9ef18e1325385082.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a980749c6b2a18c80426dd189e5506334343ca4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a980749c6b2a18c80426dd189e5506334343ca4.hip
deleted file mode 100644
index 40a903ca6124d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8a980749c6b2a18c80426dd189e5506334343ca4.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8adbdcd28cb2f078f89adf9aad2b3d4a0a477823.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8adbdcd28cb2f078f89adf9aad2b3d4a0a477823.hip
deleted file mode 100644
index 4dde718429f9e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8adbdcd28cb2f078f89adf9aad2b3d4a0a477823.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b17c082f249649eca733a8f0cdf9a1205c3e3d7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b17c082f249649eca733a8f0cdf9a1205c3e3d7.hip
deleted file mode 100644
index fde18547e502b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b17c082f249649eca733a8f0cdf9a1205c3e3d7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b9043572cabb65435627a3faf23b18d039bbcd8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b9043572cabb65435627a3faf23b18d039bbcd8.hip
deleted file mode 100644
index ea134058ed3c4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b9043572cabb65435627a3faf23b18d039bbcd8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b92990df507e82f96eeb7aa3ec00c01437566fb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b92990df507e82f96eeb7aa3ec00c01437566fb.hip
deleted file mode 100644
index 613b337ac3af9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8b92990df507e82f96eeb7aa3ec00c01437566fb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8bd1a40b12ce927323594fcce61eb9c20cc5e3d4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8bd1a40b12ce927323594fcce61eb9c20cc5e3d4.hip
deleted file mode 100644
index 45da6271b9456..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8bd1a40b12ce927323594fcce61eb9c20cc5e3d4.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8bd7b8c63a51c8639b3cf27ad09d41ae47c480d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8bd7b8c63a51c8639b3cf27ad09d41ae47c480d3.hip
deleted file mode 100644
index e9e547b584b27..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8bd7b8c63a51c8639b3cf27ad09d41ae47c480d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c074afcf33e3f3534ac3577484237fcfd2ca48e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c074afcf33e3f3534ac3577484237fcfd2ca48e.hip
deleted file mode 100644
index 76c854d485a90..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c074afcf33e3f3534ac3577484237fcfd2ca48e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c13c4f3f645a2bb475eb1c55ce1de452f0e2332.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c13c4f3f645a2bb475eb1c55ce1de452f0e2332.hip
deleted file mode 100644
index 71115563f0281..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c13c4f3f645a2bb475eb1c55ce1de452f0e2332.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c3bd4e029bba76ebfc79e6522dbc8ca0bba5dd2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c3bd4e029bba76ebfc79e6522dbc8ca0bba5dd2.hip
deleted file mode 100644
index 0bc60f2df1260..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c3bd4e029bba76ebfc79e6522dbc8ca0bba5dd2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c4688cbd23727dd0ea9a36fb977b31aeae98d65.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c4688cbd23727dd0ea9a36fb977b31aeae98d65.hip
deleted file mode 100644
index 2d098e3171b08..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c4688cbd23727dd0ea9a36fb977b31aeae98d65.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c7970957024de050748d3e31cef434f582d968b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c7970957024de050748d3e31cef434f582d968b.hip
deleted file mode 100644
index 40e2545786a43..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8c7970957024de050748d3e31cef434f582d968b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8cdcdeb845e7bcdb89ef70ab2a97157d4db3cb52.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8cdcdeb845e7bcdb89ef70ab2a97157d4db3cb52.hip
deleted file mode 100644
index ae75482000105..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8cdcdeb845e7bcdb89ef70ab2a97157d4db3cb52.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8cf1007430da272174d3476d042f398627e83512.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8cf1007430da272174d3476d042f398627e83512.hip
deleted file mode 100644
index 8e0f9e699cb16..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8cf1007430da272174d3476d042f398627e83512.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d079c1eb36db8461fa8b861c56760afcd97cc34.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d079c1eb36db8461fa8b861c56760afcd97cc34.hip
deleted file mode 100644
index dfa92cb08266f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d079c1eb36db8461fa8b861c56760afcd97cc34.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d7549e66ef309e32779ddc2a1f14e79bae53754.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d7549e66ef309e32779ddc2a1f14e79bae53754.hip
deleted file mode 100644
index 3232ed4849593..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d7549e66ef309e32779ddc2a1f14e79bae53754.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d79fe8a600c3b4e0ec9aa510f8036ba2b608985.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d79fe8a600c3b4e0ec9aa510f8036ba2b608985.hip
deleted file mode 100644
index d2c743d56bb2e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8d79fe8a600c3b4e0ec9aa510f8036ba2b608985.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8da8285bd6182355e3164cdc5a983375cdf0a61d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8da8285bd6182355e3164cdc5a983375cdf0a61d.hip
deleted file mode 100644
index a5566ff44ac4f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8da8285bd6182355e3164cdc5a983375cdf0a61d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e1b48a28b71c7f4c78eb14321b39951a7c5e903.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e1b48a28b71c7f4c78eb14321b39951a7c5e903.hip
deleted file mode 100644
index 4e55049aeccc9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e1b48a28b71c7f4c78eb14321b39951a7c5e903.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e2c587db8bd9f1b551624e0cf8b67a90245d7da.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e2c587db8bd9f1b551624e0cf8b67a90245d7da.hip
deleted file mode 100644
index 8e27eebc36c8d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e2c587db8bd9f1b551624e0cf8b67a90245d7da.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e2d5f979fc4fbd0991581a020a414f9c8656ae2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e2d5f979fc4fbd0991581a020a414f9c8656ae2.hip
deleted file mode 100644
index 06d937a7ddd5e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e2d5f979fc4fbd0991581a020a414f9c8656ae2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e431313fe082958d31b68d2fd0d61df0fe56736.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e431313fe082958d31b68d2fd0d61df0fe56736.hip
deleted file mode 100644
index 7941750c33300..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e431313fe082958d31b68d2fd0d61df0fe56736.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e50ea8dd480012cbe10be392cd26d1870e6ef9b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e50ea8dd480012cbe10be392cd26d1870e6ef9b.hip
deleted file mode 100644
index a7bf9d7c2dc1a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e50ea8dd480012cbe10be392cd26d1870e6ef9b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e675919a6c7758cbbeecb83b7ac6c62f95cdb46.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e675919a6c7758cbbeecb83b7ac6c62f95cdb46.hip
deleted file mode 100644
index 9ebe18a6d3554..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e675919a6c7758cbbeecb83b7ac6c62f95cdb46.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e812705ae3e452810794fa7caceef2ef6066dfb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e812705ae3e452810794fa7caceef2ef6066dfb.hip
deleted file mode 100644
index 4a6e1dc46efd0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e812705ae3e452810794fa7caceef2ef6066dfb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e816fcad5e9ecfca94a6491eb2274bcc41e558b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e816fcad5e9ecfca94a6491eb2274bcc41e558b.hip
deleted file mode 100644
index d78735b25f67a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e816fcad5e9ecfca94a6491eb2274bcc41e558b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e938d0e3ad30db201880642e57758285b2ec4cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e938d0e3ad30db201880642e57758285b2ec4cb.hip
deleted file mode 100644
index e3803929cc396..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8e938d0e3ad30db201880642e57758285b2ec4cb.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::fp16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8efb5fc2ace6839eac741c5e6616665845f43566.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8efb5fc2ace6839eac741c5e6616665845f43566.hip
deleted file mode 100644
index 2dd0c6ae4b214..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8efb5fc2ace6839eac741c5e6616665845f43566.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f607ee20c0d92b6dbd0338f139517fdcce98d0c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f607ee20c0d92b6dbd0338f139517fdcce98d0c.hip
deleted file mode 100644
index 8993d88485539..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f607ee20c0d92b6dbd0338f139517fdcce98d0c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f6e463eedd3e65b9c79feed3cd92ad8cbc9f036.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f6e463eedd3e65b9c79feed3cd92ad8cbc9f036.hip
deleted file mode 100644
index 574533a38ffe5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f6e463eedd3e65b9c79feed3cd92ad8cbc9f036.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f7166d4bb0c1c9b9999ba16a1adbf09ebfdb6f1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f7166d4bb0c1c9b9999ba16a1adbf09ebfdb6f1.hip
deleted file mode 100644
index a95ada1b80905..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8f7166d4bb0c1c9b9999ba16a1adbf09ebfdb6f1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fa4c40e244b412a07933d369704bcdaa6d5e74c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fa4c40e244b412a07933d369704bcdaa6d5e74c.hip
deleted file mode 100644
index e739f83a24efc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fa4c40e244b412a07933d369704bcdaa6d5e74c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fb224b40a7be7db0a9c5c08cc5ab05b526c14e8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fb224b40a7be7db0a9c5c08cc5ab05b526c14e8.hip
deleted file mode 100644
index 7553a56903c40..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fb224b40a7be7db0a9c5c08cc5ab05b526c14e8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fb33fc20f2e85e915f1b1529ae87981dfcaf86d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fb33fc20f2e85e915f1b1529ae87981dfcaf86d.hip
deleted file mode 100644
index 2feaeb0eb37e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fb33fc20f2e85e915f1b1529ae87981dfcaf86d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fc08b4f3959a2375ac03f40c4ce12d70cdc2d80.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fc08b4f3959a2375ac03f40c4ce12d70cdc2d80.hip
deleted file mode 100644
index bc051fb5d67ab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_8fc08b4f3959a2375ac03f40c4ce12d70cdc2d80.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9009b7d39346537aa6c4a4e46b81139f603edb60.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9009b7d39346537aa6c4a4e46b81139f603edb60.hip
deleted file mode 100644
index 0a983cd9f0769..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9009b7d39346537aa6c4a4e46b81139f603edb60.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_900d7f81c73b35ea64095d01c5d48d9190839e0a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_900d7f81c73b35ea64095d01c5d48d9190839e0a.hip
deleted file mode 100644
index 7a43253061cf5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_900d7f81c73b35ea64095d01c5d48d9190839e0a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9068ba8df8b0e977e9769f6acf6cfee6b00b9922.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9068ba8df8b0e977e9769f6acf6cfee6b00b9922.hip
deleted file mode 100644
index 943b25bae0285..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9068ba8df8b0e977e9769f6acf6cfee6b00b9922.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_906fa8bf5e992ddc25815486ae9c24d8bfba7227.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_906fa8bf5e992ddc25815486ae9c24d8bfba7227.hip
deleted file mode 100644
index 259d2ed9379b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_906fa8bf5e992ddc25815486ae9c24d8bfba7227.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90b17d8cba28cceddb3ef907df878aeef0762d15.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90b17d8cba28cceddb3ef907df878aeef0762d15.hip
deleted file mode 100644
index 2b3c375441f40..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90b17d8cba28cceddb3ef907df878aeef0762d15.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90da0d469cca5c8481504148468460c85a15c559.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90da0d469cca5c8481504148468460c85a15c559.hip
deleted file mode 100644
index 349f9825962f5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90da0d469cca5c8481504148468460c85a15c559.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90e5c56e92712d00092ba102a5eb5176a3e5d471.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90e5c56e92712d00092ba102a5eb5176a3e5d471.hip
deleted file mode 100644
index 45579a587006a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_90e5c56e92712d00092ba102a5eb5176a3e5d471.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_910cb8bd09d287a1566265eb1e8894fe68d3cc81.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_910cb8bd09d287a1566265eb1e8894fe68d3cc81.hip
deleted file mode 100644
index 934419fd20a7b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_910cb8bd09d287a1566265eb1e8894fe68d3cc81.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_915b75db795dbef037b14b003ee073665fe35d3e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_915b75db795dbef037b14b003ee073665fe35d3e.hip
deleted file mode 100644
index 472dcad3aba79..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_915b75db795dbef037b14b003ee073665fe35d3e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9163ae070075f26926a86d39e15c27e6edb1f1cf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9163ae070075f26926a86d39e15c27e6edb1f1cf.hip
deleted file mode 100644
index 7ba85fd26fc5f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9163ae070075f26926a86d39e15c27e6edb1f1cf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91695dea4171747fb3cc6d910459f800608d07c1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91695dea4171747fb3cc6d910459f800608d07c1.hip
deleted file mode 100644
index 6f1a50eaf8566..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91695dea4171747fb3cc6d910459f800608d07c1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_919ae177b7a793fa352c4f6bb8e4175f3064d814.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_919ae177b7a793fa352c4f6bb8e4175f3064d814.hip
deleted file mode 100644
index 319a938f0092e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_919ae177b7a793fa352c4f6bb8e4175f3064d814.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91a6200e36944b1f11106c02f7fcee053f01ee71.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91a6200e36944b1f11106c02f7fcee053f01ee71.hip
deleted file mode 100644
index b9d3651bc47eb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91a6200e36944b1f11106c02f7fcee053f01ee71.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91b9e2616c2fe0480096b1ccf0f74d584b220146.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91b9e2616c2fe0480096b1ccf0f74d584b220146.hip
deleted file mode 100644
index 64d6f0a659bb2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91b9e2616c2fe0480096b1ccf0f74d584b220146.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91c916e14198f6d18dc89915e379b01070434e91.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91c916e14198f6d18dc89915e379b01070434e91.hip
deleted file mode 100644
index 7458f3ff3d257..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_91c916e14198f6d18dc89915e379b01070434e91.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9207a63fc55c411c73e4f93306c5ffed800dd249.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9207a63fc55c411c73e4f93306c5ffed800dd249.hip
deleted file mode 100644
index 30096f88e790e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9207a63fc55c411c73e4f93306c5ffed800dd249.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92121fd448b4640a17e1a7fe73bb7b58714c0afb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92121fd448b4640a17e1a7fe73bb7b58714c0afb.hip
deleted file mode 100644
index fe0620251d5f3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92121fd448b4640a17e1a7fe73bb7b58714c0afb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_921f789d619db6f225e8e9d646e93bbc9dc1a669.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_921f789d619db6f225e8e9d646e93bbc9dc1a669.hip
deleted file mode 100644
index 60b820477d09b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_921f789d619db6f225e8e9d646e93bbc9dc1a669.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92739f4464512feee083b875e11e11eee4f5b448.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92739f4464512feee083b875e11e11eee4f5b448.hip
deleted file mode 100644
index 7f77ad653accb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92739f4464512feee083b875e11e11eee4f5b448.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92992be6252f2afdc368bd4baec4b8a55ae0abf8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92992be6252f2afdc368bd4baec4b8a55ae0abf8.hip
deleted file mode 100644
index 32e51b40be0b9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92992be6252f2afdc368bd4baec4b8a55ae0abf8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92b0770fe64e3c60b9e56170aa88bbf74802a813.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92b0770fe64e3c60b9e56170aa88bbf74802a813.hip
deleted file mode 100644
index 6acc4ea1078fa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92b0770fe64e3c60b9e56170aa88bbf74802a813.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92b722cdabcfaa388ccc6ccceb7e42462f3bdcd1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92b722cdabcfaa388ccc6ccceb7e42462f3bdcd1.hip
deleted file mode 100644
index b182fd41a86e5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92b722cdabcfaa388ccc6ccceb7e42462f3bdcd1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92ba64cdf615c1be2865f027a293cb530fc07dc6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92ba64cdf615c1be2865f027a293cb530fc07dc6.hip
deleted file mode 100644
index b2d0f026a38fd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92ba64cdf615c1be2865f027a293cb530fc07dc6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92d841e6d783bb46d841aafd9027f92dd1b61b88.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92d841e6d783bb46d841aafd9027f92dd1b61b88.hip
deleted file mode 100644
index f90fd93ffce92..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92d841e6d783bb46d841aafd9027f92dd1b61b88.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92e53359c69bbe4d7405d45261a8a62008eb7d06.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92e53359c69bbe4d7405d45261a8a62008eb7d06.hip
deleted file mode 100644
index 1f76859fac803..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92e53359c69bbe4d7405d45261a8a62008eb7d06.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92f9ad0fb65638cfffb3e7786f2cbf01d9585b23.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92f9ad0fb65638cfffb3e7786f2cbf01d9585b23.hip
deleted file mode 100644
index 343d71f5d5f7e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_92f9ad0fb65638cfffb3e7786f2cbf01d9585b23.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93054acb8a9508fd0f0f486367fb62454de47c39.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93054acb8a9508fd0f0f486367fb62454de47c39.hip
deleted file mode 100644
index cbbd56c6c3350..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93054acb8a9508fd0f0f486367fb62454de47c39.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_931cf8d05cfa45319f4e5bb49334d35a530bffcf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_931cf8d05cfa45319f4e5bb49334d35a530bffcf.hip
deleted file mode 100644
index c5626028e95e6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_931cf8d05cfa45319f4e5bb49334d35a530bffcf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93728d999ae43ee1b5a16e60b90cf8533c7d303f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93728d999ae43ee1b5a16e60b90cf8533c7d303f.hip
deleted file mode 100644
index 1ce05165d6648..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93728d999ae43ee1b5a16e60b90cf8533c7d303f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_937801fbb43fb6797f0425f08d13926b74d87c4a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_937801fbb43fb6797f0425f08d13926b74d87c4a.hip
deleted file mode 100644
index fb710a2ba7133..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_937801fbb43fb6797f0425f08d13926b74d87c4a.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_937c48d0b7096ad6c8bc445f13f2c8c1934695ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_937c48d0b7096ad6c8bc445f13f2c8c1934695ab.hip
deleted file mode 100644
index 3456274dfce7c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_937c48d0b7096ad6c8bc445f13f2c8c1934695ab.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93b885d6869400b0dc2ef1b2c2636ddfd21cde31.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93b885d6869400b0dc2ef1b2c2636ddfd21cde31.hip
deleted file mode 100644
index a93a653733373..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_93b885d6869400b0dc2ef1b2c2636ddfd21cde31.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_942439e4f5644a3a4630481bc7d98834b29b6e1c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_942439e4f5644a3a4630481bc7d98834b29b6e1c.hip
deleted file mode 100644
index f95269dbae6aa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_942439e4f5644a3a4630481bc7d98834b29b6e1c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94a94d145e575747c8956ac703810582c819e2e8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94a94d145e575747c8956ac703810582c819e2e8.hip
deleted file mode 100644
index f095f167e5a11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94a94d145e575747c8956ac703810582c819e2e8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94aa519eb57e5797125728492d9330f5c0f0670a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94aa519eb57e5797125728492d9330f5c0f0670a.hip
deleted file mode 100644
index d5b99a089e94d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94aa519eb57e5797125728492d9330f5c0f0670a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94f6f9dee9f0c3825d91f4d320a5280070e60ee7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94f6f9dee9f0c3825d91f4d320a5280070e60ee7.hip
deleted file mode 100644
index 33dbc942249ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_94f6f9dee9f0c3825d91f4d320a5280070e60ee7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_95061acc6650fc7b79fa1fe5b2b1e083555eec2c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_95061acc6650fc7b79fa1fe5b2b1e083555eec2c.hip
deleted file mode 100644
index a42f11ace2239..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_95061acc6650fc7b79fa1fe5b2b1e083555eec2c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_951343832a5bfd060c8d12da0d8a090f070a717d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_951343832a5bfd060c8d12da0d8a090f070a717d.hip
deleted file mode 100644
index 3805805858e2e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_951343832a5bfd060c8d12da0d8a090f070a717d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9545f95c1093c60f0fb6c794636f79aaeb53b733.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9545f95c1093c60f0fb6c794636f79aaeb53b733.hip
deleted file mode 100644
index 9ea99eae30f0f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9545f95c1093c60f0fb6c794636f79aaeb53b733.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_95530399ad7b43d8ce2c89da24c71056f2146b18.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_95530399ad7b43d8ce2c89da24c71056f2146b18.hip
deleted file mode 100644
index 4d406596d68ca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_95530399ad7b43d8ce2c89da24c71056f2146b18.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9583148fd684a7e6a312127e023798278415bd27.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9583148fd684a7e6a312127e023798278415bd27.hip
deleted file mode 100644
index 490b196957e67..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9583148fd684a7e6a312127e023798278415bd27.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9594816877815bc0294610ca24f986fdccdc7c6f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9594816877815bc0294610ca24f986fdccdc7c6f.hip
deleted file mode 100644
index 0253bcf4a3c30..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9594816877815bc0294610ca24f986fdccdc7c6f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_960ecb3013071fb65f2d5ed4c947c4bf303e5308.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_960ecb3013071fb65f2d5ed4c947c4bf303e5308.hip
deleted file mode 100644
index bd228e1332962..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_960ecb3013071fb65f2d5ed4c947c4bf303e5308.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9638c9618dbf2af119e37596f7eb0fd3f8d72748.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9638c9618dbf2af119e37596f7eb0fd3f8d72748.hip
deleted file mode 100644
index 0b0de06043d68..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9638c9618dbf2af119e37596f7eb0fd3f8d72748.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_963986150adcd6e1d3886bacf2166de1252e14df.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_963986150adcd6e1d3886bacf2166de1252e14df.hip
deleted file mode 100644
index 24e465ac22cf0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_963986150adcd6e1d3886bacf2166de1252e14df.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_964f916d3484295b5918e2e4c22c5529588a5662.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_964f916d3484295b5918e2e4c22c5529588a5662.hip
deleted file mode 100644
index c78ed25690b78..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_964f916d3484295b5918e2e4c22c5529588a5662.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9689ecd7bf51bcffe9f5002959bdda41c50a3c8b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9689ecd7bf51bcffe9f5002959bdda41c50a3c8b.hip
deleted file mode 100644
index 24a1c2d5bc178..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9689ecd7bf51bcffe9f5002959bdda41c50a3c8b.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_968fc75a7d102aca068e3ceb6111728c280fa837.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_968fc75a7d102aca068e3ceb6111728c280fa837.hip
deleted file mode 100644
index f85d1366ac328..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_968fc75a7d102aca068e3ceb6111728c280fa837.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96c129dd4c798343d6f78ab78056f0faf2f1c9d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96c129dd4c798343d6f78ab78056f0faf2f1c9d3.hip
deleted file mode 100644
index 83b5d416c2469..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96c129dd4c798343d6f78ab78056f0faf2f1c9d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96c5e79f54b71677124f555b0ae4bfd27248d099.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96c5e79f54b71677124f555b0ae4bfd27248d099.hip
deleted file mode 100644
index 4d3fe46ed60e4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96c5e79f54b71677124f555b0ae4bfd27248d099.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96caa2056d99eb67ada498e287b4fae984397691.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96caa2056d99eb67ada498e287b4fae984397691.hip
deleted file mode 100644
index d014ecfc46f47..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96caa2056d99eb67ada498e287b4fae984397691.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96dee49ec6755006d67f0c30c65f50558bba69b0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96dee49ec6755006d67f0c30c65f50558bba69b0.hip
deleted file mode 100644
index 684f8b4637cfe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96dee49ec6755006d67f0c30c65f50558bba69b0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96f1bb85dff8c97846f6b2e8796a6289bcd0d9d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96f1bb85dff8c97846f6b2e8796a6289bcd0d9d3.hip
deleted file mode 100644
index 223e3e97cac15..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_96f1bb85dff8c97846f6b2e8796a6289bcd0d9d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_970073c70133ff2ee4737f803a0ac43801c47242.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_970073c70133ff2ee4737f803a0ac43801c47242.hip
deleted file mode 100644
index 2b525655af45d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_970073c70133ff2ee4737f803a0ac43801c47242.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_971a08c2e48d805b295d979b24173a04cf58def0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_971a08c2e48d805b295d979b24173a04cf58def0.hip
deleted file mode 100644
index b8d6364662294..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_971a08c2e48d805b295d979b24173a04cf58def0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_97246460c21bc66c0f13936d27477a9fca1c44d1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_97246460c21bc66c0f13936d27477a9fca1c44d1.hip
deleted file mode 100644
index e9da11dde247e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_97246460c21bc66c0f13936d27477a9fca1c44d1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9745b04a8026a01828c5dd606d89d044d3ed1d99.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9745b04a8026a01828c5dd606d89d044d3ed1d99.hip
deleted file mode 100644
index 2b7c4276f99cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9745b04a8026a01828c5dd606d89d044d3ed1d99.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_976cf509d9c2bf86ba6ee5ded544fa8e6717f590.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_976cf509d9c2bf86ba6ee5ded544fa8e6717f590.hip
deleted file mode 100644
index e56905bd29a92..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_976cf509d9c2bf86ba6ee5ded544fa8e6717f590.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_977137b371df841993c8d0584be7d83aca6add78.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_977137b371df841993c8d0584be7d83aca6add78.hip
deleted file mode 100644
index b8418a6d53511..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_977137b371df841993c8d0584be7d83aca6add78.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_97851d5ecbf02f8af623988b1a39c0b91e51533a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_97851d5ecbf02f8af623988b1a39c0b91e51533a.hip
deleted file mode 100644
index 044cd6b8ed273..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_97851d5ecbf02f8af623988b1a39c0b91e51533a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9801b25e0f132d647934deb395b62a3f70cc7c88.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9801b25e0f132d647934deb395b62a3f70cc7c88.hip
deleted file mode 100644
index e3b12b6278796..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9801b25e0f132d647934deb395b62a3f70cc7c88.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_987a617fae00fa90a1ba60937b0312c81087c19e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_987a617fae00fa90a1ba60937b0312c81087c19e.hip
deleted file mode 100644
index 4cd907c270147..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_987a617fae00fa90a1ba60937b0312c81087c19e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_987f00dd759d9714693e7517dfaa8bb427294d42.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_987f00dd759d9714693e7517dfaa8bb427294d42.hip
deleted file mode 100644
index 05e2906409fa0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_987f00dd759d9714693e7517dfaa8bb427294d42.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9893336a4b00b2a63f23ed7e13ec54c82d9e5063.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9893336a4b00b2a63f23ed7e13ec54c82d9e5063.hip
deleted file mode 100644
index 2c14aa3dfe800..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9893336a4b00b2a63f23ed7e13ec54c82d9e5063.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98e484adeddf3394d8d7693b808d83b64c71ee69.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98e484adeddf3394d8d7693b808d83b64c71ee69.hip
deleted file mode 100644
index 73f1520faa2ba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98e484adeddf3394d8d7693b808d83b64c71ee69.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98f5efcd500ce6b9ffc14bc9877e0ba457539925.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98f5efcd500ce6b9ffc14bc9877e0ba457539925.hip
deleted file mode 100644
index ea866c827c29e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98f5efcd500ce6b9ffc14bc9877e0ba457539925.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98f9a4f4d85f292b78123599a2e1798f12aa545b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98f9a4f4d85f292b78123599a2e1798f12aa545b.hip
deleted file mode 100644
index da636b8e344fe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_98f9a4f4d85f292b78123599a2e1798f12aa545b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9990e6ad243a48b84304b5cad0c663c0802aedfd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9990e6ad243a48b84304b5cad0c663c0802aedfd.hip
deleted file mode 100644
index 3be0fb6df1ac7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9990e6ad243a48b84304b5cad0c663c0802aedfd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99ae680eed89ea93a3a94586bd5a68dbc5439f37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99ae680eed89ea93a3a94586bd5a68dbc5439f37.hip
deleted file mode 100644
index 420137406e6da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99ae680eed89ea93a3a94586bd5a68dbc5439f37.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99e2f290b962f1617b0a9d4fd6d55c43e4439d6f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99e2f290b962f1617b0a9d4fd6d55c43e4439d6f.hip
deleted file mode 100644
index 145cf83ea4bdf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99e2f290b962f1617b0a9d4fd6d55c43e4439d6f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99f8352674bd6bbe98944a1c0a769a4fc028a623.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99f8352674bd6bbe98944a1c0a769a4fc028a623.hip
deleted file mode 100644
index ee4361de8718d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_99f8352674bd6bbe98944a1c0a769a4fc028a623.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a0a70932bd587759df1e5e150b25b0126d7b529.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a0a70932bd587759df1e5e150b25b0126d7b529.hip
deleted file mode 100644
index af0ffa2137788..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a0a70932bd587759df1e5e150b25b0126d7b529.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a20fa19d8d30654602e363806f559113218d66d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a20fa19d8d30654602e363806f559113218d66d.hip
deleted file mode 100644
index c9a76092cd766..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a20fa19d8d30654602e363806f559113218d66d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a8e04fe9432a60f86ff0369e8c1851821074a04.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a8e04fe9432a60f86ff0369e8c1851821074a04.hip
deleted file mode 100644
index d596d50779bbd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a8e04fe9432a60f86ff0369e8c1851821074a04.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a9edbe35a8fac7796f00bde836bd547044770ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a9edbe35a8fac7796f00bde836bd547044770ea.hip
deleted file mode 100644
index e5c9c1d2d93d7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9a9edbe35a8fac7796f00bde836bd547044770ea.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ab73ea77ec20ea3bfaf995dacf93a6960ecdca0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ab73ea77ec20ea3bfaf995dacf93a6960ecdca0.hip
deleted file mode 100644
index e973f986deb3a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ab73ea77ec20ea3bfaf995dacf93a6960ecdca0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ad1f99284aafc8d7908d062f179a056eb314925.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ad1f99284aafc8d7908d062f179a056eb314925.hip
deleted file mode 100644
index 95bf476babcbf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ad1f99284aafc8d7908d062f179a056eb314925.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ae866c7db36286876818bfb718ac35204fa3843.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ae866c7db36286876818bfb718ac35204fa3843.hip
deleted file mode 100644
index a15c188623361..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ae866c7db36286876818bfb718ac35204fa3843.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9afe4b6f3b901ff4af81bd4f1cd8ff19f09d0b07.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9afe4b6f3b901ff4af81bd4f1cd8ff19f09d0b07.hip
deleted file mode 100644
index d0e5277b6f850..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9afe4b6f3b901ff4af81bd4f1cd8ff19f09d0b07.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b062dd633645772e4f2caffd111af73184f7657.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b062dd633645772e4f2caffd111af73184f7657.hip
deleted file mode 100644
index dc6a112cc5f24..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b062dd633645772e4f2caffd111af73184f7657.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b327f0fa1155f2235d76be45cd22e3db5a69429.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b327f0fa1155f2235d76be45cd22e3db5a69429.hip
deleted file mode 100644
index 064c9fcbd8fac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b327f0fa1155f2235d76be45cd22e3db5a69429.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b4dcde1ae3446b825dea739d4295c1d1ec5c4be.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b4dcde1ae3446b825dea739d4295c1d1ec5c4be.hip
deleted file mode 100644
index d566a2583a3c6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b4dcde1ae3446b825dea739d4295c1d1ec5c4be.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b6d08e63b9a90f2524cbfa8c5fcf8b82a1d2d36.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b6d08e63b9a90f2524cbfa8c5fcf8b82a1d2d36.hip
deleted file mode 100644
index 50f7dfc8afb3d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b6d08e63b9a90f2524cbfa8c5fcf8b82a1d2d36.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b73c92a13757877f34bd8a13c6fb29b60999020.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b73c92a13757877f34bd8a13c6fb29b60999020.hip
deleted file mode 100644
index 1ccea0d2aec36..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b73c92a13757877f34bd8a13c6fb29b60999020.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b841b7cf5da31f0c30ec42c91cc8d5bd3fedd03.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b841b7cf5da31f0c30ec42c91cc8d5bd3fedd03.hip
deleted file mode 100644
index 3eb3770ae4233..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9b841b7cf5da31f0c30ec42c91cc8d5bd3fedd03.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9bcc791049e3ff9ebc1a9085d2d20efcc2f99b71.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9bcc791049e3ff9ebc1a9085d2d20efcc2f99b71.hip
deleted file mode 100644
index aec0230d741be..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9bcc791049e3ff9ebc1a9085d2d20efcc2f99b71.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9bf235679af1ca03a6e601b4cf6cd0416d1c9091.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9bf235679af1ca03a6e601b4cf6cd0416d1c9091.hip
deleted file mode 100644
index 82c5b3f6119d4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9bf235679af1ca03a6e601b4cf6cd0416d1c9091.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9c4fc7cda4b560040cec93f63021b529aa1ee3fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9c4fc7cda4b560040cec93f63021b529aa1ee3fd.hip
deleted file mode 100644
index 9cd1460c12a50..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9c4fc7cda4b560040cec93f63021b529aa1ee3fd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ca3b1d36d777213eb381b47871bf15dd163c994.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ca3b1d36d777213eb381b47871bf15dd163c994.hip
deleted file mode 100644
index c99e8d61b1777..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9ca3b1d36d777213eb381b47871bf15dd163c994.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9cc3ef3d3b36f52089548e9dce522b0448e2c26a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9cc3ef3d3b36f52089548e9dce522b0448e2c26a.hip
deleted file mode 100644
index ee1287b43f36b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9cc3ef3d3b36f52089548e9dce522b0448e2c26a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d3d274058bc0a3d4d35d90669587761fdfbdba1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d3d274058bc0a3d4d35d90669587761fdfbdba1.hip
deleted file mode 100644
index 1f64d97348235..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d3d274058bc0a3d4d35d90669587761fdfbdba1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d6759d8855c4c6289f1f241a1628cf0406c1b64.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d6759d8855c4c6289f1f241a1628cf0406c1b64.hip
deleted file mode 100644
index 588cc32456d7f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d6759d8855c4c6289f1f241a1628cf0406c1b64.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d69d441f48f9ea346dd8e00376a9a708da3ad87.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d69d441f48f9ea346dd8e00376a9a708da3ad87.hip
deleted file mode 100644
index 11a8723145a99..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9d69d441f48f9ea346dd8e00376a9a708da3ad87.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9dc424f0e192155e3c4e786e5b87d5a1a3e6c4ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9dc424f0e192155e3c4e786e5b87d5a1a3e6c4ad.hip
deleted file mode 100644
index 2cbd253a36a8c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9dc424f0e192155e3c4e786e5b87d5a1a3e6c4ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9e51083e13aa4dfa8c969f8f916835a8e5e9ca39.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9e51083e13aa4dfa8c969f8f916835a8e5e9ca39.hip
deleted file mode 100644
index 3cfc1ba422ed6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9e51083e13aa4dfa8c969f8f916835a8e5e9ca39.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9eef1b54d5d3841f3fa6b84cca6c7ad33efa2d9f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9eef1b54d5d3841f3fa6b84cca6c7ad33efa2d9f.hip
deleted file mode 100644
index 3b4f0f7ad35ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9eef1b54d5d3841f3fa6b84cca6c7ad33efa2d9f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9f0517550c7a23882b95de451e8099ea2186b4ce.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9f0517550c7a23882b95de451e8099ea2186b4ce.hip
deleted file mode 100644
index 4d63d6f0b9355..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9f0517550c7a23882b95de451e8099ea2186b4ce.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9fb389d4b5ba590baa951f17da06f0e53d2bfa55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9fb389d4b5ba590baa951f17da06f0e53d2bfa55.hip
deleted file mode 100644
index 3f741d34a0fbb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_9fb389d4b5ba590baa951f17da06f0e53d2bfa55.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a017be7b8bcf303b30a147f41346898acc5fab7d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a017be7b8bcf303b30a147f41346898acc5fab7d.hip
deleted file mode 100644
index ffd897040dd74..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a017be7b8bcf303b30a147f41346898acc5fab7d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a02a71fdd587e47ee68e0cc76c3c4494ce06c359.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a02a71fdd587e47ee68e0cc76c3c4494ce06c359.hip
deleted file mode 100644
index 7ca210ee7e1fe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a02a71fdd587e47ee68e0cc76c3c4494ce06c359.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a02f152e9184af0b3d77082d8bdf519dbbfceb2d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a02f152e9184af0b3d77082d8bdf519dbbfceb2d.hip
deleted file mode 100644
index e4aa5e9c9a9dd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a02f152e9184af0b3d77082d8bdf519dbbfceb2d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a046e888e3836b0bd3c49fec8e1872e880798f0c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a046e888e3836b0bd3c49fec8e1872e880798f0c.hip
deleted file mode 100644
index 371c3c4660736..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a046e888e3836b0bd3c49fec8e1872e880798f0c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a0874fc5ac87a1ec487c7722bf3b1bdaa924ee09.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a0874fc5ac87a1ec487c7722bf3b1bdaa924ee09.hip
deleted file mode 100644
index 4524055af0e3d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a0874fc5ac87a1ec487c7722bf3b1bdaa924ee09.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a094599fb5caf5e7aba728cd4713a8d0c6368a46.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a094599fb5caf5e7aba728cd4713a8d0c6368a46.hip
deleted file mode 100644
index 5802a84a0e0bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a094599fb5caf5e7aba728cd4713a8d0c6368a46.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a0a556c9358ddd6db719458c81d2d6d822a895da.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a0a556c9358ddd6db719458c81d2d6d822a895da.hip
deleted file mode 100644
index 68adac38978a5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a0a556c9358ddd6db719458c81d2d6d822a895da.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a103cd47156a98ad2cf2c325ea00df3f1d67fb72.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a103cd47156a98ad2cf2c325ea00df3f1d67fb72.hip
deleted file mode 100644
index 673ecb9ada48d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a103cd47156a98ad2cf2c325ea00df3f1d67fb72.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a189292c81a18d21a2921ce6740f81ebf4c046ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a189292c81a18d21a2921ce6740f81ebf4c046ad.hip
deleted file mode 100644
index 7214c13980e8d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a189292c81a18d21a2921ce6740f81ebf4c046ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1c71e7d33f0597fe090a3524e33e18b2e562680.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1c71e7d33f0597fe090a3524e33e18b2e562680.hip
deleted file mode 100644
index 110d3ba13feac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1c71e7d33f0597fe090a3524e33e18b2e562680.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1cba1509c413c870c5d784410855ee1bd737da2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1cba1509c413c870c5d784410855ee1bd737da2.hip
deleted file mode 100644
index 0955208d8ab0f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1cba1509c413c870c5d784410855ee1bd737da2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1d6ad9de7ac7993ae1923a2ef070b7dacb8c563.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1d6ad9de7ac7993ae1923a2ef070b7dacb8c563.hip
deleted file mode 100644
index af8d021fa7bed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a1d6ad9de7ac7993ae1923a2ef070b7dacb8c563.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a20c91b2f11bb7e5058ca7935b0bda4f5558a9dc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a20c91b2f11bb7e5058ca7935b0bda4f5558a9dc.hip
deleted file mode 100644
index 3f80acb260ea6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a20c91b2f11bb7e5058ca7935b0bda4f5558a9dc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a21f3637624762547af1292e1b85e640b1d329dc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a21f3637624762547af1292e1b85e640b1d329dc.hip
deleted file mode 100644
index d26979e2d5d39..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a21f3637624762547af1292e1b85e640b1d329dc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a225c4f1f3c7b271957768bb9235131c67afb48a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a225c4f1f3c7b271957768bb9235131c67afb48a.hip
deleted file mode 100644
index 061afd23dc9f7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a225c4f1f3c7b271957768bb9235131c67afb48a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2482a64659c838f3da55f56e3cbbee1dbfe6722.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2482a64659c838f3da55f56e3cbbee1dbfe6722.hip
deleted file mode 100644
index 3321e8a41a54f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2482a64659c838f3da55f56e3cbbee1dbfe6722.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a25e2aed617e1ff31f93ae7e054313ee0dceee97.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a25e2aed617e1ff31f93ae7e054313ee0dceee97.hip
deleted file mode 100644
index 592870480fd8d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a25e2aed617e1ff31f93ae7e054313ee0dceee97.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2a715b7e9c1a576f011dfe5769c5b392e984f82.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2a715b7e9c1a576f011dfe5769c5b392e984f82.hip
deleted file mode 100644
index bb05144138887..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2a715b7e9c1a576f011dfe5769c5b392e984f82.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2ef5d30a2318ae06430d17f84878800c4ca7364.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2ef5d30a2318ae06430d17f84878800c4ca7364.hip
deleted file mode 100644
index 566198c7f81e2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a2ef5d30a2318ae06430d17f84878800c4ca7364.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3339150d8bf9d073827738527f6cbe15b854607.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3339150d8bf9d073827738527f6cbe15b854607.hip
deleted file mode 100644
index 21c141e4088bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3339150d8bf9d073827738527f6cbe15b854607.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3709e4fc53d2254a03ea7660b8c72d2f47cf1ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3709e4fc53d2254a03ea7660b8c72d2f47cf1ad.hip
deleted file mode 100644
index e46b8bbff6865..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3709e4fc53d2254a03ea7660b8c72d2f47cf1ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a388a284f45f711d82a6ed87036d87cef1872eb1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a388a284f45f711d82a6ed87036d87cef1872eb1.hip
deleted file mode 100644
index 7c7b7b8746cf6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a388a284f45f711d82a6ed87036d87cef1872eb1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3ac4f93722dc314086f1b7d7b8adc687cd75f82.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3ac4f93722dc314086f1b7d7b8adc687cd75f82.hip
deleted file mode 100644
index 1d0892ee5bd93..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3ac4f93722dc314086f1b7d7b8adc687cd75f82.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3d7aa46528ee74e2bef1e87c1feceacfa55e173.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3d7aa46528ee74e2bef1e87c1feceacfa55e173.hip
deleted file mode 100644
index a1a71baf2510c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3d7aa46528ee74e2bef1e87c1feceacfa55e173.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3dc780b17152f696f9b957432c2eae8fb16e85e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3dc780b17152f696f9b957432c2eae8fb16e85e.hip
deleted file mode 100644
index 780d9b66c8349..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3dc780b17152f696f9b957432c2eae8fb16e85e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3f9c236d24b30bc9c3fad90cfd6eb00da835de2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3f9c236d24b30bc9c3fad90cfd6eb00da835de2.hip
deleted file mode 100644
index 7bb86590805b1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3f9c236d24b30bc9c3fad90cfd6eb00da835de2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3ff8445ba691807caadd9f26e7eb90851875280.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3ff8445ba691807caadd9f26e7eb90851875280.hip
deleted file mode 100644
index 8964c12872662..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a3ff8445ba691807caadd9f26e7eb90851875280.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a421c2ed6b295c458071f1988b9d6f7b46e8992c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a421c2ed6b295c458071f1988b9d6f7b46e8992c.hip
deleted file mode 100644
index 55057543df912..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a421c2ed6b295c458071f1988b9d6f7b46e8992c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4700d87a19a173e84d64e43cffabbed52366e35.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4700d87a19a173e84d64e43cffabbed52366e35.hip
deleted file mode 100644
index b3775081f0aae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4700d87a19a173e84d64e43cffabbed52366e35.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a487f617c4b84c6a0328fedac750d41dc3dafe27.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a487f617c4b84c6a0328fedac750d41dc3dafe27.hip
deleted file mode 100644
index 0fe779c20a88b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a487f617c4b84c6a0328fedac750d41dc3dafe27.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a48843d844f78690c7a45b730652f0f763c595c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a48843d844f78690c7a45b730652f0f763c595c7.hip
deleted file mode 100644
index e762047290efd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a48843d844f78690c7a45b730652f0f763c595c7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4980becb0d3149fee575bad1fc3b463d08aabf5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4980becb0d3149fee575bad1fc3b463d08aabf5.hip
deleted file mode 100644
index 25037c1333b1b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4980becb0d3149fee575bad1fc3b463d08aabf5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4b7f10440331a8a88ff93ba253217c2832bcf9e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4b7f10440331a8a88ff93ba253217c2832bcf9e.hip
deleted file mode 100644
index 7813101645422..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a4b7f10440331a8a88ff93ba253217c2832bcf9e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a55b47aafc4340e69e300ac61a7601a5c14513b7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a55b47aafc4340e69e300ac61a7601a5c14513b7.hip
deleted file mode 100644
index df41b38772432..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a55b47aafc4340e69e300ac61a7601a5c14513b7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a55c7dd576e5b1061c059e5e99aeedf4389e2d25.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a55c7dd576e5b1061c059e5e99aeedf4389e2d25.hip
deleted file mode 100644
index a830e93c2d460..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a55c7dd576e5b1061c059e5e99aeedf4389e2d25.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a59423c095db052603d77073d409534bceef425f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a59423c095db052603d77073d409534bceef425f.hip
deleted file mode 100644
index e51048b4fc902..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a59423c095db052603d77073d409534bceef425f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5a7833f4597bb03a3e845d5580d677e97421040.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5a7833f4597bb03a3e845d5580d677e97421040.hip
deleted file mode 100644
index fc3fea65dc686..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5a7833f4597bb03a3e845d5580d677e97421040.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5bdc110955c05c6c6ea236a6f60266a4a6dce5e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5bdc110955c05c6c6ea236a6f60266a4a6dce5e.hip
deleted file mode 100644
index fe03517034b76..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5bdc110955c05c6c6ea236a6f60266a4a6dce5e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5c0109313de1f6245d2a80f8539485b849e9d55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5c0109313de1f6245d2a80f8539485b849e9d55.hip
deleted file mode 100644
index 7e8564220a37d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5c0109313de1f6245d2a80f8539485b849e9d55.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5c4dc0d70c547dbbfb661e879ba7f9adfafc2ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5c4dc0d70c547dbbfb661e879ba7f9adfafc2ea.hip
deleted file mode 100644
index acd7584593b03..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5c4dc0d70c547dbbfb661e879ba7f9adfafc2ea.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5d4eb673bafd81e3a0ee213da4603d88b8460ec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5d4eb673bafd81e3a0ee213da4603d88b8460ec.hip
deleted file mode 100644
index 3bf976b5b2e80..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5d4eb673bafd81e3a0ee213da4603d88b8460ec.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5e5cae764142683b70d3344cf07dd1edb7d69e2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5e5cae764142683b70d3344cf07dd1edb7d69e2.hip
deleted file mode 100644
index bcc20d9d38328..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5e5cae764142683b70d3344cf07dd1edb7d69e2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5f2f0cef657ae5e333d65ae4ab20529a43cd7de.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5f2f0cef657ae5e333d65ae4ab20529a43cd7de.hip
deleted file mode 100644
index 8788ff8255384..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5f2f0cef657ae5e333d65ae4ab20529a43cd7de.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5f8b7b2a891aa9f2ab49762eb31d835efdf18b6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5f8b7b2a891aa9f2ab49762eb31d835efdf18b6.hip
deleted file mode 100644
index f72f0eb7cf7d9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5f8b7b2a891aa9f2ab49762eb31d835efdf18b6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5fa94bb32a80e81886b711ebfcf2df5f5405866.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5fa94bb32a80e81886b711ebfcf2df5f5405866.hip
deleted file mode 100644
index c04f2790cb063..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a5fa94bb32a80e81886b711ebfcf2df5f5405866.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a622fa57764ec746e02f6d4bd4846b48c722b807.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a622fa57764ec746e02f6d4bd4846b48c722b807.hip
deleted file mode 100644
index 59dc4d1674c2e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a622fa57764ec746e02f6d4bd4846b48c722b807.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a62a2ab489839ea1a1bfd1b24e54a3c232ed934f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a62a2ab489839ea1a1bfd1b24e54a3c232ed934f.hip
deleted file mode 100644
index 7f13ed4266c5a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a62a2ab489839ea1a1bfd1b24e54a3c232ed934f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a6461d72fb6ba50e81de3f661528c96dcfdc3f3c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a6461d72fb6ba50e81de3f661528c96dcfdc3f3c.hip
deleted file mode 100644
index 0037c2b61515c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a6461d72fb6ba50e81de3f661528c96dcfdc3f3c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a64b4cf3f6706e4b4e0af4402e2263b9a1585f9b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a64b4cf3f6706e4b4e0af4402e2263b9a1585f9b.hip
deleted file mode 100644
index bdcd8611d9cb2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a64b4cf3f6706e4b4e0af4402e2263b9a1585f9b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a65c43b870705c780d734f9ef063f55cf8b3b52d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a65c43b870705c780d734f9ef063f55cf8b3b52d.hip
deleted file mode 100644
index af5d24650029b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a65c43b870705c780d734f9ef063f55cf8b3b52d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a673f35edd69241c6b921d6712dfd064d78ecbad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a673f35edd69241c6b921d6712dfd064d78ecbad.hip
deleted file mode 100644
index a8bf1bced5bbb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a673f35edd69241c6b921d6712dfd064d78ecbad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a71305f191f06cd53b7563971c706e8b71b19e2f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a71305f191f06cd53b7563971c706e8b71b19e2f.hip
deleted file mode 100644
index 54c9e1be56154..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a71305f191f06cd53b7563971c706e8b71b19e2f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a74b0e7dd816ad08eec5a1bba6e227afee9813ec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a74b0e7dd816ad08eec5a1bba6e227afee9813ec.hip
deleted file mode 100644
index cb6605dd510f2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a74b0e7dd816ad08eec5a1bba6e227afee9813ec.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a7784b03ad757d51c234fa86ea9891f055ecd5c1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a7784b03ad757d51c234fa86ea9891f055ecd5c1.hip
deleted file mode 100644
index 7e4ac2125a720..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a7784b03ad757d51c234fa86ea9891f055ecd5c1.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a78fecb9725ceb4bcf2aa037d43bc43efeb1c3fd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a78fecb9725ceb4bcf2aa037d43bc43efeb1c3fd.hip
deleted file mode 100644
index 829a327af50df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a78fecb9725ceb4bcf2aa037d43bc43efeb1c3fd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a7f7553a7d2f6d42fe695cdc64423c85223af440.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a7f7553a7d2f6d42fe695cdc64423c85223af440.hip
deleted file mode 100644
index f35e0dd453d73..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a7f7553a7d2f6d42fe695cdc64423c85223af440.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a821661d8280c6e9d27f2c9ce1b3c855387b5a76.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a821661d8280c6e9d27f2c9ce1b3c855387b5a76.hip
deleted file mode 100644
index 271d757689322..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a821661d8280c6e9d27f2c9ce1b3c855387b5a76.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a85d35b2fd98742427930eb536e346ffb005edd8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a85d35b2fd98742427930eb536e346ffb005edd8.hip
deleted file mode 100644
index 61b95a2c1454e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a85d35b2fd98742427930eb536e346ffb005edd8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a8a4af070ee46d802cb11086b93daf91538f8a04.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a8a4af070ee46d802cb11086b93daf91538f8a04.hip
deleted file mode 100644
index 2fccdb851edf1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a8a4af070ee46d802cb11086b93daf91538f8a04.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a8a744edfa3a19d1493611df5bd0d4d59b707d43.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a8a744edfa3a19d1493611df5bd0d4d59b707d43.hip
deleted file mode 100644
index 1b0e1a78094f4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a8a744edfa3a19d1493611df5bd0d4d59b707d43.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a92b43d374642df991edef1f6036dc898bf77cf8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a92b43d374642df991edef1f6036dc898bf77cf8.hip
deleted file mode 100644
index 578fdbdb463d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a92b43d374642df991edef1f6036dc898bf77cf8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a93324ccf11b273ed20fd960c61df897c8890b1d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a93324ccf11b273ed20fd960c61df897c8890b1d.hip
deleted file mode 100644
index ad910a4ad9505..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a93324ccf11b273ed20fd960c61df897c8890b1d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a93a03b33305b33055273711ab31a5b8d8298d5d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a93a03b33305b33055273711ab31a5b8d8298d5d.hip
deleted file mode 100644
index 4127c092ffbd6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a93a03b33305b33055273711ab31a5b8d8298d5d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a968df29f5ae1463706b7981b3bde55918e1aa65.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a968df29f5ae1463706b7981b3bde55918e1aa65.hip
deleted file mode 100644
index b5ab69da5f612..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a968df29f5ae1463706b7981b3bde55918e1aa65.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a98925d99dc484da41dd55700e151cf545cf821d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a98925d99dc484da41dd55700e151cf545cf821d.hip
deleted file mode 100644
index 5966147a0b5f4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a98925d99dc484da41dd55700e151cf545cf821d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9b50c6ebb27986ce5b378d8c39315eb9cb91dea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9b50c6ebb27986ce5b378d8c39315eb9cb91dea.hip
deleted file mode 100644
index dbbdbaaf80ba3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9b50c6ebb27986ce5b378d8c39315eb9cb91dea.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9d2be18e2d53a5144f97dfdebb225fcb6d611d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9d2be18e2d53a5144f97dfdebb225fcb6d611d3.hip
deleted file mode 100644
index 8d7cdebbd730a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9d2be18e2d53a5144f97dfdebb225fcb6d611d3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9df9ac4ee78e5f4d5bd0567e58a7090907c61e1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9df9ac4ee78e5f4d5bd0567e58a7090907c61e1.hip
deleted file mode 100644
index f91d5eac3fa59..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9df9ac4ee78e5f4d5bd0567e58a7090907c61e1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9f00f270680de81df7737e848e0408cb070e68b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9f00f270680de81df7737e848e0408cb070e68b.hip
deleted file mode 100644
index 73efa28b3162e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_a9f00f270680de81df7737e848e0408cb070e68b.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, false, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa1041530f794c7b8dc4a8321ea0fcdd338fff35.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa1041530f794c7b8dc4a8321ea0fcdd338fff35.hip
deleted file mode 100644
index f384c617e3ea5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa1041530f794c7b8dc4a8321ea0fcdd338fff35.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa522b43c5e5ea69bcabb4c0fe28def2bd081a12.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa522b43c5e5ea69bcabb4c0fe28def2bd081a12.hip
deleted file mode 100644
index 328a9c14353e2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa522b43c5e5ea69bcabb4c0fe28def2bd081a12.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa6d13b09f85ee62bb5018608812181fb43afc86.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa6d13b09f85ee62bb5018608812181fb43afc86.hip
deleted file mode 100644
index 55cd7f9899919..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa6d13b09f85ee62bb5018608812181fb43afc86.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa82d20635e592edbf00439294835f6f39ad54a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa82d20635e592edbf00439294835f6f39ad54a3.hip
deleted file mode 100644
index cef0937d502ca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa82d20635e592edbf00439294835f6f39ad54a3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa996b9c843200a2ec33ed4319b48106cd7c6384.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa996b9c843200a2ec33ed4319b48106cd7c6384.hip
deleted file mode 100644
index 9a23fd3a853a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aa996b9c843200a2ec33ed4319b48106cd7c6384.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aafe891dad43815e635f81225705ff944f990d75.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aafe891dad43815e635f81225705ff944f990d75.hip
deleted file mode 100644
index 655ee83d8e968..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aafe891dad43815e635f81225705ff944f990d75.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab09941bddfa9d61985b55f9b6bf0edec9bb89f6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab09941bddfa9d61985b55f9b6bf0edec9bb89f6.hip
deleted file mode 100644
index d9b5508c76d9e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab09941bddfa9d61985b55f9b6bf0edec9bb89f6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab0be5a2072b5e87f5ee58149688796b6513219f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab0be5a2072b5e87f5ee58149688796b6513219f.hip
deleted file mode 100644
index 586302a0693aa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab0be5a2072b5e87f5ee58149688796b6513219f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab0c3fe9529e24327686070731d0ac3ada76245e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab0c3fe9529e24327686070731d0ac3ada76245e.hip
deleted file mode 100644
index 2f9ca5ec1d7aa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab0c3fe9529e24327686070731d0ac3ada76245e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab1ca4ce061f7f69a250356f613cab00d1e2ac71.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab1ca4ce061f7f69a250356f613cab00d1e2ac71.hip
deleted file mode 100644
index ee98cd327b7a6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab1ca4ce061f7f69a250356f613cab00d1e2ac71.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab1d7f93427095e39bfc1d986b3d7fe54073ec75.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab1d7f93427095e39bfc1d986b3d7fe54073ec75.hip
deleted file mode 100644
index cd3f04da24ecd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab1d7f93427095e39bfc1d986b3d7fe54073ec75.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab43f4a56c166dad0113f51b337a083f4df7cdb6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab43f4a56c166dad0113f51b337a083f4df7cdb6.hip
deleted file mode 100644
index 8a79d192d5471..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab43f4a56c166dad0113f51b337a083f4df7cdb6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab56e886d53a1d88fada0f10f00b9f398dc54568.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab56e886d53a1d88fada0f10f00b9f398dc54568.hip
deleted file mode 100644
index 3eadd1bd28e5d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab56e886d53a1d88fada0f10f00b9f398dc54568.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab6cd5c9242f8278c8f3d9ce57b97d605c7e5a3e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab6cd5c9242f8278c8f3d9ce57b97d605c7e5a3e.hip
deleted file mode 100644
index e7d75ae4cd2da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab6cd5c9242f8278c8f3d9ce57b97d605c7e5a3e.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab877ae2a1aab04498bf2b26b3fe99d6488ef151.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab877ae2a1aab04498bf2b26b3fe99d6488ef151.hip
deleted file mode 100644
index 9734f97bd608f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ab877ae2a1aab04498bf2b26b3fe99d6488ef151.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_abf6c6412f9853855b74a96e862935ddef66f763.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_abf6c6412f9853855b74a96e862935ddef66f763.hip
deleted file mode 100644
index afa30674780f9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_abf6c6412f9853855b74a96e862935ddef66f763.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_abf92a5314fd33491b5eb6ebd2418b7e0d5db774.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_abf92a5314fd33491b5eb6ebd2418b7e0d5db774.hip
deleted file mode 100644
index 22742e24afbeb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_abf92a5314fd33491b5eb6ebd2418b7e0d5db774.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac1ccde31b47e0e56ee0daab6403fed7895208c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac1ccde31b47e0e56ee0daab6403fed7895208c7.hip
deleted file mode 100644
index 5600e3189fbd3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac1ccde31b47e0e56ee0daab6403fed7895208c7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac5e9aee85cd16903bf7b82a4ac10402b0b26e22.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac5e9aee85cd16903bf7b82a4ac10402b0b26e22.hip
deleted file mode 100644
index e2cf68e6382ac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac5e9aee85cd16903bf7b82a4ac10402b0b26e22.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac9382cf8bb56ffd962c99329bf67da992f8810d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac9382cf8bb56ffd962c99329bf67da992f8810d.hip
deleted file mode 100644
index f7c405f44d871..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ac9382cf8bb56ffd962c99329bf67da992f8810d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aceb0641213e9a45ba48bcf72bb23845720d8b79.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aceb0641213e9a45ba48bcf72bb23845720d8b79.hip
deleted file mode 100644
index 5d50b3f96963e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aceb0641213e9a45ba48bcf72bb23845720d8b79.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad091c69d19b27f7ad50ef6311532ad8b642a9c6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad091c69d19b27f7ad50ef6311532ad8b642a9c6.hip
deleted file mode 100644
index c45321177eabf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad091c69d19b27f7ad50ef6311532ad8b642a9c6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad82071cc074fd30437f6158b5eb2c6df1f8c587.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad82071cc074fd30437f6158b5eb2c6df1f8c587.hip
deleted file mode 100644
index 2a409b3c11917..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad82071cc074fd30437f6158b5eb2c6df1f8c587.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad989d2ce769f20e175fa88f4082c1c25fe03062.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad989d2ce769f20e175fa88f4082c1c25fe03062.hip
deleted file mode 100644
index 2e5b2e2ffda9d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad989d2ce769f20e175fa88f4082c1c25fe03062.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad9b99a194b59d3149842c15733394da275b12c0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad9b99a194b59d3149842c15733394da275b12c0.hip
deleted file mode 100644
index bd65652c940dc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ad9b99a194b59d3149842c15733394da275b12c0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ada016be2bd0e377fbe01fa7adb9bbb8febce100.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ada016be2bd0e377fbe01fa7adb9bbb8febce100.hip
deleted file mode 100644
index b70e9584be9b2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ada016be2bd0e377fbe01fa7adb9bbb8febce100.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adae2d4f8b2dac799e03ea6f279e6ecdf66f5381.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adae2d4f8b2dac799e03ea6f279e6ecdf66f5381.hip
deleted file mode 100644
index 27b5bb9e27eaa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adae2d4f8b2dac799e03ea6f279e6ecdf66f5381.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adaef10ff2c5d89530310bdf1d53a194f06a94ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adaef10ff2c5d89530310bdf1d53a194f06a94ef.hip
deleted file mode 100644
index b842623af7352..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adaef10ff2c5d89530310bdf1d53a194f06a94ef.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_add29e3e9828911a117dccaa5650e77805730d14.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_add29e3e9828911a117dccaa5650e77805730d14.hip
deleted file mode 100644
index 71ed5a679714a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_add29e3e9828911a117dccaa5650e77805730d14.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adda7ad787524e3e47dcc1b65c41b2faea38f55f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adda7ad787524e3e47dcc1b65c41b2faea38f55f.hip
deleted file mode 100644
index 092b0e389fed3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adda7ad787524e3e47dcc1b65c41b2faea38f55f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_addb6a14043c5a4df0f5042b3770b40c4e90795c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_addb6a14043c5a4df0f5042b3770b40c4e90795c.hip
deleted file mode 100644
index 396e283283b72..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_addb6a14043c5a4df0f5042b3770b40c4e90795c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adf160741a4f751d2f15d6eb23d4121cdca62b55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adf160741a4f751d2f15d6eb23d4121cdca62b55.hip
deleted file mode 100644
index 961d81dcbdc8b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_adf160741a4f751d2f15d6eb23d4121cdca62b55.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae1ab1f4bbe86bb9bbc22e4774648076c321136f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae1ab1f4bbe86bb9bbc22e4774648076c321136f.hip
deleted file mode 100644
index 675ad03969cdc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae1ab1f4bbe86bb9bbc22e4774648076c321136f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae1afeb6cfdf860ff08e4c2f11c922fd5bfa621a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae1afeb6cfdf860ff08e4c2f11c922fd5bfa621a.hip
deleted file mode 100644
index 2a1126b12c369..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae1afeb6cfdf860ff08e4c2f11c922fd5bfa621a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae239476d61f48379754b97f29d7a285cc3192de.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae239476d61f48379754b97f29d7a285cc3192de.hip
deleted file mode 100644
index 20f5842fd656b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae239476d61f48379754b97f29d7a285cc3192de.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae4e7253ad4873576052ec0a9400597bb7975753.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae4e7253ad4873576052ec0a9400597bb7975753.hip
deleted file mode 100644
index 40395e507c4f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae4e7253ad4873576052ec0a9400597bb7975753.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae4e80cb185759dd9b3eb3c67c239964b3694caa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae4e80cb185759dd9b3eb3c67c239964b3694caa.hip
deleted file mode 100644
index 7c97156107877..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae4e80cb185759dd9b3eb3c67c239964b3694caa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae51b30c7e1cd30e550187458350c8db7c59a9ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae51b30c7e1cd30e550187458350c8db7c59a9ef.hip
deleted file mode 100644
index 9aa4cef1e17c6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae51b30c7e1cd30e550187458350c8db7c59a9ef.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae7899b1ef159ecbf01f27014601eb79b31b49b3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae7899b1ef159ecbf01f27014601eb79b31b49b3.hip
deleted file mode 100644
index 4b42db971d1d2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae7899b1ef159ecbf01f27014601eb79b31b49b3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae87b1d5c50606430b544ed650d87df24366e7d5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae87b1d5c50606430b544ed650d87df24366e7d5.hip
deleted file mode 100644
index bf947e99de0f3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae87b1d5c50606430b544ed650d87df24366e7d5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae8d0bdde763e617beafc0365ec4a3cd11df6c55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae8d0bdde763e617beafc0365ec4a3cd11df6c55.hip
deleted file mode 100644
index 2b3cf7a363bfd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ae8d0bdde763e617beafc0365ec4a3cd11df6c55.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebb2441e6cc1ccba4a391566e547402bcf7ced2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebb2441e6cc1ccba4a391566e547402bcf7ced2.hip
deleted file mode 100644
index b996d43df67b1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebb2441e6cc1ccba4a391566e547402bcf7ced2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebd5fed34ebceb879ae3dffaf58c7c04ab5fe80.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebd5fed34ebceb879ae3dffaf58c7c04ab5fe80.hip
deleted file mode 100644
index 63771416681fd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebd5fed34ebceb879ae3dffaf58c7c04ab5fe80.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebff7e6605b273bad844b8f70ef031625bff48e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebff7e6605b273bad844b8f70ef031625bff48e.hip
deleted file mode 100644
index 18aa32f29a1b5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aebff7e6605b273bad844b8f70ef031625bff48e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aec87e65afa93e84d7a947c52f291c1c7360033c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aec87e65afa93e84d7a947c52f291c1c7360033c.hip
deleted file mode 100644
index a2e054816ec78..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aec87e65afa93e84d7a947c52f291c1c7360033c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aece14f7a220222eb4ce6783ec2b9fce6fde94b8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aece14f7a220222eb4ce6783ec2b9fce6fde94b8.hip
deleted file mode 100644
index b65dffe783fda..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_aece14f7a220222eb4ce6783ec2b9fce6fde94b8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_af06c0dae15684f83e15722a4c07342af9ea011c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_af06c0dae15684f83e15722a4c07342af9ea011c.hip
deleted file mode 100644
index d419f27d1be2a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_af06c0dae15684f83e15722a4c07342af9ea011c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_af6ccfa11add1ae49888337e84d9c446d2f67da4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_af6ccfa11add1ae49888337e84d9c446d2f67da4.hip
deleted file mode 100644
index 494f7e4bde698..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_af6ccfa11add1ae49888337e84d9c446d2f67da4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afadc4f76e237514db0bc0203102297b79730bd0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afadc4f76e237514db0bc0203102297b79730bd0.hip
deleted file mode 100644
index fbdf5cdc2a44c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afadc4f76e237514db0bc0203102297b79730bd0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afc4b47a6fa62a4ca5cff6a7e01c9f6b371d2215.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afc4b47a6fa62a4ca5cff6a7e01c9f6b371d2215.hip
deleted file mode 100644
index d2858c19b9eb2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afc4b47a6fa62a4ca5cff6a7e01c9f6b371d2215.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afcafd07c1f56e74373ccf37db35976023456d50.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afcafd07c1f56e74373ccf37db35976023456d50.hip
deleted file mode 100644
index 3432aa47b55db..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afcafd07c1f56e74373ccf37db35976023456d50.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afccf699f593c828e11efc053b144044e45b32d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afccf699f593c828e11efc053b144044e45b32d6.hip
deleted file mode 100644
index e32ac0ef383e4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afccf699f593c828e11efc053b144044e45b32d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afda8f46b5ded4c2aa9d722fec17b75004b59f7d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afda8f46b5ded4c2aa9d722fec17b75004b59f7d.hip
deleted file mode 100644
index 32d191fbc037e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afda8f46b5ded4c2aa9d722fec17b75004b59f7d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afdab954fd111ec48721f25710d61c0c8affd8db.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afdab954fd111ec48721f25710d61c0c8affd8db.hip
deleted file mode 100644
index f631f97586686..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_afdab954fd111ec48721f25710d61c0c8affd8db.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b00e062055933388e37525df5766f3c14cd3538a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b00e062055933388e37525df5766f3c14cd3538a.hip
deleted file mode 100644
index f6ca9c9f53039..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b00e062055933388e37525df5766f3c14cd3538a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b01dc872c24db4db0c9179fc07e17f41060390de.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b01dc872c24db4db0c9179fc07e17f41060390de.hip
deleted file mode 100644
index 479d807564524..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b01dc872c24db4db0c9179fc07e17f41060390de.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b03ab68e33844f97aa58d463e00037bc11c50da0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b03ab68e33844f97aa58d463e00037bc11c50da0.hip
deleted file mode 100644
index fb955f754811b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b03ab68e33844f97aa58d463e00037bc11c50da0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b04f14f829eff73afaa57a875f74ebd1e6860979.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b04f14f829eff73afaa57a875f74ebd1e6860979.hip
deleted file mode 100644
index a4c291be568f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b04f14f829eff73afaa57a875f74ebd1e6860979.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0544a38dfdf4d81dc95894387845f48435e299a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0544a38dfdf4d81dc95894387845f48435e299a.hip
deleted file mode 100644
index d25407ebe7288..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0544a38dfdf4d81dc95894387845f48435e299a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0dd965d5d9080ed5c6a04b7eea9890f3a264f20.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0dd965d5d9080ed5c6a04b7eea9890f3a264f20.hip
deleted file mode 100644
index 15c3c16d28f78..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0dd965d5d9080ed5c6a04b7eea9890f3a264f20.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0f555b74ed36f1bef8f47880b3edc6760f27788.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0f555b74ed36f1bef8f47880b3edc6760f27788.hip
deleted file mode 100644
index d419a0e79df21..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b0f555b74ed36f1bef8f47880b3edc6760f27788.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1766695dbb790bd614b83dc7569ad449404cc89.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1766695dbb790bd614b83dc7569ad449404cc89.hip
deleted file mode 100644
index 5baf6bfebba0e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1766695dbb790bd614b83dc7569ad449404cc89.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b18a615e66d7cd739ce35412811359a03cb23a8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b18a615e66d7cd739ce35412811359a03cb23a8e.hip
deleted file mode 100644
index e61d86ba0f11c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b18a615e66d7cd739ce35412811359a03cb23a8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b192c55f002d8540d5f965cc4df0c2e33f4b9ff9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b192c55f002d8540d5f965cc4df0c2e33f4b9ff9.hip
deleted file mode 100644
index 1d57e0d31c22f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b192c55f002d8540d5f965cc4df0c2e33f4b9ff9.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    64,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<64, ck_tile::bf16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b19f05f6848403480ba41d37cdbf44ccca1b1f8d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b19f05f6848403480ba41d37cdbf44ccca1b1f8d.hip
deleted file mode 100644
index c46c86bc35ea5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b19f05f6848403480ba41d37cdbf44ccca1b1f8d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1ad101ce91348266d3885afdf2996a0fdb72135.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1ad101ce91348266d3885afdf2996a0fdb72135.hip
deleted file mode 100644
index adb9abaa08663..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1ad101ce91348266d3885afdf2996a0fdb72135.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1c5d55d47d6038e9162d32ac968ff58c0942938.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1c5d55d47d6038e9162d32ac968ff58c0942938.hip
deleted file mode 100644
index 2e9ec621eff02..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b1c5d55d47d6038e9162d32ac968ff58c0942938.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b20c6252863a73341b0010191fad4c834860f884.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b20c6252863a73341b0010191fad4c834860f884.hip
deleted file mode 100644
index 6f4bf9142f38e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b20c6252863a73341b0010191fad4c834860f884.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b20e314642cf565e4f32bceffdb5c0e653ab627b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b20e314642cf565e4f32bceffdb5c0e653ab627b.hip
deleted file mode 100644
index 32ffb75dcb99e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b20e314642cf565e4f32bceffdb5c0e653ab627b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b24f91dec2029b25d0d96962528410df55a468ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b24f91dec2029b25d0d96962528410df55a468ed.hip
deleted file mode 100644
index 5f679a4c38877..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b24f91dec2029b25d0d96962528410df55a468ed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b285e2f1970b78e18002464eeda63798229bbc3a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b285e2f1970b78e18002464eeda63798229bbc3a.hip
deleted file mode 100644
index c90705b7782d4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b285e2f1970b78e18002464eeda63798229bbc3a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b298e213f927b518c693660110f08bdd94990ef0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b298e213f927b518c693660110f08bdd94990ef0.hip
deleted file mode 100644
index 3ec87d74291dd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b298e213f927b518c693660110f08bdd94990ef0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b2af5f5b5ee3ae964824a3e9c7bbeb5bb39c557c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b2af5f5b5ee3ae964824a3e9c7bbeb5bb39c557c.hip
deleted file mode 100644
index fcc5641cf9e10..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b2af5f5b5ee3ae964824a3e9c7bbeb5bb39c557c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b2f91e937b427ecc932c0cb0c90b2c2378db0be6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b2f91e937b427ecc932c0cb0c90b2c2378db0be6.hip
deleted file mode 100644
index 4c74ec252cfd6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b2f91e937b427ecc932c0cb0c90b2c2378db0be6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3063d06723ac70c5f8802ab49c5c35e1debf56e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3063d06723ac70c5f8802ab49c5c35e1debf56e.hip
deleted file mode 100644
index de835dc163e7f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3063d06723ac70c5f8802ab49c5c35e1debf56e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b31f56244076c501cb09b4b90975132cae4c4386.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b31f56244076c501cb09b4b90975132cae4c4386.hip
deleted file mode 100644
index 23b0e206766bf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b31f56244076c501cb09b4b90975132cae4c4386.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3486244e0b7d6dbcaa1951e8b8883ce441c3f99.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3486244e0b7d6dbcaa1951e8b8883ce441c3f99.hip
deleted file mode 100644
index dcd7724f05340..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3486244e0b7d6dbcaa1951e8b8883ce441c3f99.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b34c1ce348c3d9cdf6bbec9758de9d5fe94c43fc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b34c1ce348c3d9cdf6bbec9758de9d5fe94c43fc.hip
deleted file mode 100644
index 4736e14de6eab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b34c1ce348c3d9cdf6bbec9758de9d5fe94c43fc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b38a1d3cffae01332a3a9d9472ff1b2c443e82af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b38a1d3cffae01332a3a9d9472ff1b2c443e82af.hip
deleted file mode 100644
index f76a1230b4851..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b38a1d3cffae01332a3a9d9472ff1b2c443e82af.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3a104733f678193068d8642d6560faa03897258.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3a104733f678193068d8642d6560faa03897258.hip
deleted file mode 100644
index 619619939d991..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3a104733f678193068d8642d6560faa03897258.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3da22d3482738a8474ae15e8e5fca9020c4e195.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3da22d3482738a8474ae15e8e5fca9020c4e195.hip
deleted file mode 100644
index 3cacc0f7f1305..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b3da22d3482738a8474ae15e8e5fca9020c4e195.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41735d250b5a16967281a5f07873b9cde3df4d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41735d250b5a16967281a5f07873b9cde3df4d6.hip
deleted file mode 100644
index 104c9ad00c8b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41735d250b5a16967281a5f07873b9cde3df4d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41a30092e8138877c1f6c25656e0f8ae2c2444e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41a30092e8138877c1f6c25656e0f8ae2c2444e.hip
deleted file mode 100644
index 48ec2a43608be..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41a30092e8138877c1f6c25656e0f8ae2c2444e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41ea5293bc1c56efa2c4b5681d965aa6f2ce6c3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41ea5293bc1c56efa2c4b5681d965aa6f2ce6c3.hip
deleted file mode 100644
index af9fc1e76a567..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b41ea5293bc1c56efa2c4b5681d965aa6f2ce6c3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4588379eaa268d79fe8f8e4457b009f204a5fb7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4588379eaa268d79fe8f8e4457b009f204a5fb7.hip
deleted file mode 100644
index e30f149cc484c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4588379eaa268d79fe8f8e4457b009f204a5fb7.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::fp16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b493c99888d82cd2852bfb101f99a2e6a27665b8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b493c99888d82cd2852bfb101f99a2e6a27665b8.hip
deleted file mode 100644
index 5535687adb2d2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b493c99888d82cd2852bfb101f99a2e6a27665b8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4a5715b550f67b8870ba66e1e6282a26cc1dbf3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4a5715b550f67b8870ba66e1e6282a26cc1dbf3.hip
deleted file mode 100644
index d2db97ae443c0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4a5715b550f67b8870ba66e1e6282a26cc1dbf3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4b037a2e262d11d3ed7d9feeb41b9e05427a739.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4b037a2e262d11d3ed7d9feeb41b9e05427a739.hip
deleted file mode 100644
index e5aed0abffcf4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4b037a2e262d11d3ed7d9feeb41b9e05427a739.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4bd2d206ceb237ed2c51f58abb5cbf96e39d07b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4bd2d206ceb237ed2c51f58abb5cbf96e39d07b.hip
deleted file mode 100644
index 2c5b18529c017..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4bd2d206ceb237ed2c51f58abb5cbf96e39d07b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4ec377c44ac18527ca6a01bc3b146706a6e1e09.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4ec377c44ac18527ca6a01bc3b146706a6e1e09.hip
deleted file mode 100644
index 7d8eb8203a20f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4ec377c44ac18527ca6a01bc3b146706a6e1e09.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4f12f10d7b968e0d8e7c23f36d3a360de74a905.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4f12f10d7b968e0d8e7c23f36d3a360de74a905.hip
deleted file mode 100644
index 6e5334c6fe86e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b4f12f10d7b968e0d8e7c23f36d3a360de74a905.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b50e6df20a2426abd3d2ff2262a37c009196024c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b50e6df20a2426abd3d2ff2262a37c009196024c.hip
deleted file mode 100644
index bf900936535df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b50e6df20a2426abd3d2ff2262a37c009196024c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b513834918d5ea789e2db21abece7c2d3532a7e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b513834918d5ea789e2db21abece7c2d3532a7e7.hip
deleted file mode 100644
index 4d9e1ab628edc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b513834918d5ea789e2db21abece7c2d3532a7e7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5248f443a12d96815c04409a00102923c717023.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5248f443a12d96815c04409a00102923c717023.hip
deleted file mode 100644
index 3a025b2120166..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5248f443a12d96815c04409a00102923c717023.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5371415448fffffd58bf014dac9f4876153657b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5371415448fffffd58bf014dac9f4876153657b.hip
deleted file mode 100644
index c1f77b871dd8c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5371415448fffffd58bf014dac9f4876153657b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5ac596c636df55e81293228cbc53dcbb3024e5a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5ac596c636df55e81293228cbc53dcbb3024e5a.hip
deleted file mode 100644
index 81275b67f37af..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5ac596c636df55e81293228cbc53dcbb3024e5a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5ba2e73df35f6e0f7317303823fde92a42b1a35.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5ba2e73df35f6e0f7317303823fde92a42b1a35.hip
deleted file mode 100644
index c50c8a851c37a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5ba2e73df35f6e0f7317303823fde92a42b1a35.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5bccc85f74f54a2ceb17fe3040b04fe306c53f9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5bccc85f74f54a2ceb17fe3040b04fe306c53f9.hip
deleted file mode 100644
index aa77681e56f1f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5bccc85f74f54a2ceb17fe3040b04fe306c53f9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5c3131fb8e5a25bd4a14bc9075eb6fa01b61d02.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5c3131fb8e5a25bd4a14bc9075eb6fa01b61d02.hip
deleted file mode 100644
index 21bc1da79f987..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5c3131fb8e5a25bd4a14bc9075eb6fa01b61d02.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5c7fca1f76a31b0390e92d90d569fab94d4f783.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5c7fca1f76a31b0390e92d90d569fab94d4f783.hip
deleted file mode 100644
index 4b2061e07074b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5c7fca1f76a31b0390e92d90d569fab94d4f783.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5db3d5b1d8af89381fc4b8073f84c5fa25fdef5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5db3d5b1d8af89381fc4b8073f84c5fa25fdef5.hip
deleted file mode 100644
index a633d45545e64..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b5db3d5b1d8af89381fc4b8073f84c5fa25fdef5.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b60a4e87a7aabfe3c1ce02b408522f3ec862e3d7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b60a4e87a7aabfe3c1ce02b408522f3ec862e3d7.hip
deleted file mode 100644
index 4aac2ba6d11fd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b60a4e87a7aabfe3c1ce02b408522f3ec862e3d7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b6b17ae67adee9e56a022cd2a5514fb9c4e99920.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b6b17ae67adee9e56a022cd2a5514fb9c4e99920.hip
deleted file mode 100644
index b4dd9a5f26f24..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b6b17ae67adee9e56a022cd2a5514fb9c4e99920.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b72a804bb3c99830653d41ac0bd49943c801b89a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b72a804bb3c99830653d41ac0bd49943c801b89a.hip
deleted file mode 100644
index 49f0019e84d09..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b72a804bb3c99830653d41ac0bd49943c801b89a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b737410b404a51043fc3bd503c0b107c297e4c9f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b737410b404a51043fc3bd503c0b107c297e4c9f.hip
deleted file mode 100644
index 669c4f50cc2a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b737410b404a51043fc3bd503c0b107c297e4c9f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b75843bb13058ffe29251e053800c509c7590544.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b75843bb13058ffe29251e053800c509c7590544.hip
deleted file mode 100644
index 8661dd83e17d6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b75843bb13058ffe29251e053800c509c7590544.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b774450ebadaacf23e944aaf8ca90eada01e8a5a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b774450ebadaacf23e944aaf8ca90eada01e8a5a.hip
deleted file mode 100644
index cd5ca1de7620b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b774450ebadaacf23e944aaf8ca90eada01e8a5a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b779cc0b0380e1e6a2b51fc6216fdd72215b882b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b779cc0b0380e1e6a2b51fc6216fdd72215b882b.hip
deleted file mode 100644
index 57931b5693145..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b779cc0b0380e1e6a2b51fc6216fdd72215b882b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b7a03ab0b7887cc7ed0cb40e56360a8d36c0bb8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b7a03ab0b7887cc7ed0cb40e56360a8d36c0bb8e.hip
deleted file mode 100644
index 900530389684f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b7a03ab0b7887cc7ed0cb40e56360a8d36c0bb8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b80d0828ba6d24ea3c1a97bd9835ee937b4b32fb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b80d0828ba6d24ea3c1a97bd9835ee937b4b32fb.hip
deleted file mode 100644
index 3f2cf269e13ce..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b80d0828ba6d24ea3c1a97bd9835ee937b4b32fb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b872f9e6ebe330cc1818ea82b53acec79a2f672c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b872f9e6ebe330cc1818ea82b53acec79a2f672c.hip
deleted file mode 100644
index 9831aad701fdf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b872f9e6ebe330cc1818ea82b53acec79a2f672c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b8fbc6f6e9c515edce3c7a438b3bc308b30d3857.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b8fbc6f6e9c515edce3c7a438b3bc308b30d3857.hip
deleted file mode 100644
index 11c22fde0ecdb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b8fbc6f6e9c515edce3c7a438b3bc308b30d3857.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9385db12001110c42eff6aabad935a69ad3afe2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9385db12001110c42eff6aabad935a69ad3afe2.hip
deleted file mode 100644
index 1e0d64fb54a29..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9385db12001110c42eff6aabad935a69ad3afe2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9559dd36a0a4f5e068a722e285f485137bd5ef0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9559dd36a0a4f5e068a722e285f485137bd5ef0.hip
deleted file mode 100644
index 176a67e35c82b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9559dd36a0a4f5e068a722e285f485137bd5ef0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9627f9c8d0088df0364a64643f2b5dcd951f2bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9627f9c8d0088df0364a64643f2b5dcd951f2bb.hip
deleted file mode 100644
index 987ed2a7e99f6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9627f9c8d0088df0364a64643f2b5dcd951f2bb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9a742ceeb6736a2c8f9439d0b05e10d3e0c5c6f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9a742ceeb6736a2c8f9439d0b05e10d3e0c5c6f.hip
deleted file mode 100644
index 69fda307714b3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9a742ceeb6736a2c8f9439d0b05e10d3e0c5c6f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9baf70220079e6d4e87eb01a7259923d8a01e29.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9baf70220079e6d4e87eb01a7259923d8a01e29.hip
deleted file mode 100644
index d3507927033ca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9baf70220079e6d4e87eb01a7259923d8a01e29.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9d00ab8373747a5c6b9d2f8dd50ceb14db4163c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9d00ab8373747a5c6b9d2f8dd50ceb14db4163c.hip
deleted file mode 100644
index eaa319d9dede3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9d00ab8373747a5c6b9d2f8dd50ceb14db4163c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9ed0a64deb55616646ea98b21a891c971cd98ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9ed0a64deb55616646ea98b21a891c971cd98ad.hip
deleted file mode 100644
index f189b923c12b0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_b9ed0a64deb55616646ea98b21a891c971cd98ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ba145535e53899fe127987aa854f81234a9c51c4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ba145535e53899fe127987aa854f81234a9c51c4.hip
deleted file mode 100644
index a118777df1c00..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ba145535e53899fe127987aa854f81234a9c51c4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ba8b09f0aaa40a7c9ad5f0458b460d3e328f3c74.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ba8b09f0aaa40a7c9ad5f0458b460d3e328f3c74.hip
deleted file mode 100644
index c3c11ed810a18..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ba8b09f0aaa40a7c9ad5f0458b460d3e328f3c74.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bafbef3f13d429ec3e9f4672218998d5669d79f2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bafbef3f13d429ec3e9f4672218998d5669d79f2.hip
deleted file mode 100644
index 4a7bd27846622..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bafbef3f13d429ec3e9f4672218998d5669d79f2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb111b7acc269f8d5e70915d3efde4c425aa5f5c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb111b7acc269f8d5e70915d3efde4c425aa5f5c.hip
deleted file mode 100644
index 914c23091d137..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb111b7acc269f8d5e70915d3efde4c425aa5f5c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb28a4e95723e3df380f98b5ac107c4df353850b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb28a4e95723e3df380f98b5ac107c4df353850b.hip
deleted file mode 100644
index ab5a9eb00d58d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb28a4e95723e3df380f98b5ac107c4df353850b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb35c86443cc9ea38c06ebc0656306483c95ef67.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb35c86443cc9ea38c06ebc0656306483c95ef67.hip
deleted file mode 100644
index ee2462813c7b6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bb35c86443cc9ea38c06ebc0656306483c95ef67.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bba10ecb79ede07324e1198a71a95ff26e9eb235.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bba10ecb79ede07324e1198a71a95ff26e9eb235.hip
deleted file mode 100644
index 46ff62a912644..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bba10ecb79ede07324e1198a71a95ff26e9eb235.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bbe23201fbebed25781f249e5c77c31e0e7f9ddb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bbe23201fbebed25781f249e5c77c31e0e7f9ddb.hip
deleted file mode 100644
index 5c43fe5940150..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bbe23201fbebed25781f249e5c77c31e0e7f9ddb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bbfd025488e52b97c04995c4c5faff371b77e4d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bbfd025488e52b97c04995c4c5faff371b77e4d6.hip
deleted file mode 100644
index a03a3986cd5ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bbfd025488e52b97c04995c4c5faff371b77e4d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc1ae1dddb8cc5d78196da6b26ebe66c1ce7e567.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc1ae1dddb8cc5d78196da6b26ebe66c1ce7e567.hip
deleted file mode 100644
index 3d4e06bd8c450..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc1ae1dddb8cc5d78196da6b26ebe66c1ce7e567.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc238fd2095b26a167b41cdec8280182330b7b25.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc238fd2095b26a167b41cdec8280182330b7b25.hip
deleted file mode 100644
index 76feafe68d3ba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc238fd2095b26a167b41cdec8280182330b7b25.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc4425e30a0b17e8b31726817e8d3177b5c51934.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc4425e30a0b17e8b31726817e8d3177b5c51934.hip
deleted file mode 100644
index 14f0af66653c1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc4425e30a0b17e8b31726817e8d3177b5c51934.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc4e0f0496a34d2fb43c80ce0162ad4183f29064.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc4e0f0496a34d2fb43c80ce0162ad4183f29064.hip
deleted file mode 100644
index c371a0edef8ea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc4e0f0496a34d2fb43c80ce0162ad4183f29064.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc6ce17223d8d83a64b8c96ac88223e4441a4692.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc6ce17223d8d83a64b8c96ac88223e4441a4692.hip
deleted file mode 100644
index 08f3f0d96dd23..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc6ce17223d8d83a64b8c96ac88223e4441a4692.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc744db85d4237ee9640f1658e0caab7648e3bb6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc744db85d4237ee9640f1658e0caab7648e3bb6.hip
deleted file mode 100644
index d7de39eb8832b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc744db85d4237ee9640f1658e0caab7648e3bb6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc79e255d25744725e2a9db9f90d5cc2b8a0e0c1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc79e255d25744725e2a9db9f90d5cc2b8a0e0c1.hip
deleted file mode 100644
index 3e3fbb173f3d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc79e255d25744725e2a9db9f90d5cc2b8a0e0c1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc897852a4ca992961843144f4ec4f8b86dd5e9d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc897852a4ca992961843144f4ec4f8b86dd5e9d.hip
deleted file mode 100644
index 0a0ee3dfa2adc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bc897852a4ca992961843144f4ec4f8b86dd5e9d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcb6f0730fd09b4c6c60913425927dfdb8f83d82.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcb6f0730fd09b4c6c60913425927dfdb8f83d82.hip
deleted file mode 100644
index 752a14102de5c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcb6f0730fd09b4c6c60913425927dfdb8f83d82.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcd7ccdceb7baf3b986f2a0248827822a5f72e47.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcd7ccdceb7baf3b986f2a0248827822a5f72e47.hip
deleted file mode 100644
index 42deef39d3b9c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcd7ccdceb7baf3b986f2a0248827822a5f72e47.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcf8836c8cf932cc2748e313885003f0e11a887f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcf8836c8cf932cc2748e313885003f0e11a887f.hip
deleted file mode 100644
index f6540f5fd1846..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bcf8836c8cf932cc2748e313885003f0e11a887f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd064e302ff5b983dbdb4ccf51383fb29ddff44f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd064e302ff5b983dbdb4ccf51383fb29ddff44f.hip
deleted file mode 100644
index 2e015b44dfb6b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd064e302ff5b983dbdb4ccf51383fb29ddff44f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd28203f47b6a48e9b66302cf8312f3796ca500c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd28203f47b6a48e9b66302cf8312f3796ca500c.hip
deleted file mode 100644
index 128e3a8c8141c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd28203f47b6a48e9b66302cf8312f3796ca500c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd37f4f7914805a97d5073f1ebf8a8b8c2648d31.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd37f4f7914805a97d5073f1ebf8a8b8c2648d31.hip
deleted file mode 100644
index d9289c3ec2d01..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd37f4f7914805a97d5073f1ebf8a8b8c2648d31.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd3daa5f99b4522d932334924347353ce2854821.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd3daa5f99b4522d932334924347353ce2854821.hip
deleted file mode 100644
index 5ae1b0d197c8f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd3daa5f99b4522d932334924347353ce2854821.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd6aa39d0ae3c87d011610cdb5e2e317f337c454.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd6aa39d0ae3c87d011610cdb5e2e317f337c454.hip
deleted file mode 100644
index 11e91f97f28e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd6aa39d0ae3c87d011610cdb5e2e317f337c454.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd80a1774d8b7d8bee4e8663392b97cda11dcbf5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd80a1774d8b7d8bee4e8663392b97cda11dcbf5.hip
deleted file mode 100644
index 4bacc1bfcc44b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd80a1774d8b7d8bee4e8663392b97cda11dcbf5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd8bf7c572c1984ca3061062cf3c31d993f6762d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd8bf7c572c1984ca3061062cf3c31d993f6762d.hip
deleted file mode 100644
index e28aa30d4171a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd8bf7c572c1984ca3061062cf3c31d993f6762d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd9c47f3305e47db6ab6bc627fb3d80269633074.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd9c47f3305e47db6ab6bc627fb3d80269633074.hip
deleted file mode 100644
index 33e0067cef0b2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bd9c47f3305e47db6ab6bc627fb3d80269633074.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bdab172627718278a71a93e3737ef08ad9259a4f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bdab172627718278a71a93e3737ef08ad9259a4f.hip
deleted file mode 100644
index d37b4f45f3844..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bdab172627718278a71a93e3737ef08ad9259a4f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bde24a8dbe6add6f2dd2beb48b1280f3a84a9b2a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bde24a8dbe6add6f2dd2beb48b1280f3a84a9b2a.hip
deleted file mode 100644
index cda1c331b7b28..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bde24a8dbe6add6f2dd2beb48b1280f3a84a9b2a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be1e1533fc37b41838bd37edc2b6d2f2e76ae1c6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be1e1533fc37b41838bd37edc2b6d2f2e76ae1c6.hip
deleted file mode 100644
index 6ec17e72b7984..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be1e1533fc37b41838bd37edc2b6d2f2e76ae1c6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be4dd90ccb2f258029d0156cf23f940b694cf08d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be4dd90ccb2f258029d0156cf23f940b694cf08d.hip
deleted file mode 100644
index b393c952b0eeb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be4dd90ccb2f258029d0156cf23f940b694cf08d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be8ec1163a01b9cd9a802d8b44669e8770c20234.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be8ec1163a01b9cd9a802d8b44669e8770c20234.hip
deleted file mode 100644
index 52ff8d83f82ba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_be8ec1163a01b9cd9a802d8b44669e8770c20234.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_beae876d6da465687f162136231f15767cc7bb14.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_beae876d6da465687f162136231f15767cc7bb14.hip
deleted file mode 100644
index 2b5c01be4bef8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_beae876d6da465687f162136231f15767cc7bb14.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_beb9afccc15de7dfcb2e7d898abc0d61201de73e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_beb9afccc15de7dfcb2e7d898abc0d61201de73e.hip
deleted file mode 100644
index 5e1a6311a89a5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_beb9afccc15de7dfcb2e7d898abc0d61201de73e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bec30e7107c5dce3fe6aa87d83ed96da75478da0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bec30e7107c5dce3fe6aa87d83ed96da75478da0.hip
deleted file mode 100644
index 767ffc85d9533..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bec30e7107c5dce3fe6aa87d83ed96da75478da0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bec9e4c0317e8d351f60258ed6611fbf365c4024.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bec9e4c0317e8d351f60258ed6611fbf365c4024.hip
deleted file mode 100644
index 008d0a0d6e256..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bec9e4c0317e8d351f60258ed6611fbf365c4024.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_becc2a4d7ac045365300bf8bd45fc6d3e1e1c8b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_becc2a4d7ac045365300bf8bd45fc6d3e1e1c8b1.hip
deleted file mode 100644
index 8d64350df6d07..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_becc2a4d7ac045365300bf8bd45fc6d3e1e1c8b1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bed5a8c5cf683f6dfaefad72c2e2f5c2f2b2732f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bed5a8c5cf683f6dfaefad72c2e2f5c2f2b2732f.hip
deleted file mode 100644
index 9076df6359996..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bed5a8c5cf683f6dfaefad72c2e2f5c2f2b2732f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bef3bd014a918feddadc98eed92a7734f9bcd890.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bef3bd014a918feddadc98eed92a7734f9bcd890.hip
deleted file mode 100644
index 33daac294c8ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bef3bd014a918feddadc98eed92a7734f9bcd890.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bf9cdf86a7944cd690b0fcbbaec235863acd10bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bf9cdf86a7944cd690b0fcbbaec235863acd10bb.hip
deleted file mode 100644
index 191e41a09b8c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_bf9cdf86a7944cd690b0fcbbaec235863acd10bb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0338fbc05f86270ded7df2bd3e2758a03961b62.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0338fbc05f86270ded7df2bd3e2758a03961b62.hip
deleted file mode 100644
index 602016ba584ad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0338fbc05f86270ded7df2bd3e2758a03961b62.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0342686e4efd26413c6719782ed13603479c4e0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0342686e4efd26413c6719782ed13603479c4e0.hip
deleted file mode 100644
index bc752acb627f0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0342686e4efd26413c6719782ed13603479c4e0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c063318cb851ccaa923be12d34c84d839bc64bb8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c063318cb851ccaa923be12d34c84d839bc64bb8.hip
deleted file mode 100644
index 394bfbd91fd92..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c063318cb851ccaa923be12d34c84d839bc64bb8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c08095341ca7e3a1debeb780c1878e351692bee2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c08095341ca7e3a1debeb780c1878e351692bee2.hip
deleted file mode 100644
index 70eae96b196db..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c08095341ca7e3a1debeb780c1878e351692bee2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0a3c4ac0a50bb9b7ad764929dbee98c856b1210.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0a3c4ac0a50bb9b7ad764929dbee98c856b1210.hip
deleted file mode 100644
index 6b46e8921ccf6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0a3c4ac0a50bb9b7ad764929dbee98c856b1210.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0f76aff077c28f8afd7b22f284cf2894e08a043.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0f76aff077c28f8afd7b22f284cf2894e08a043.hip
deleted file mode 100644
index 27417e41b7d4f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c0f76aff077c28f8afd7b22f284cf2894e08a043.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c112c01d201c366bdd7acccf2e1b18b00f671153.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c112c01d201c366bdd7acccf2e1b18b00f671153.hip
deleted file mode 100644
index 0f5c8d6e4adb6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c112c01d201c366bdd7acccf2e1b18b00f671153.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c11d68fe766fc753c657362673704005b538660b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c11d68fe766fc753c657362673704005b538660b.hip
deleted file mode 100644
index 60f2bf6159e74..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c11d68fe766fc753c657362673704005b538660b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c137c03bf161b2ec6a9a046fa49d7bbf80ae47b8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c137c03bf161b2ec6a9a046fa49d7bbf80ae47b8.hip
deleted file mode 100644
index c48443e51c40d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c137c03bf161b2ec6a9a046fa49d7bbf80ae47b8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c197d1f050f42d82e6851fa286db6f81ba197f40.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c197d1f050f42d82e6851fa286db6f81ba197f40.hip
deleted file mode 100644
index 13a9d77fef106..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c197d1f050f42d82e6851fa286db6f81ba197f40.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1b76bc7a17f573c0d52c07ae9ff4302662ae61f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1b76bc7a17f573c0d52c07ae9ff4302662ae61f.hip
deleted file mode 100644
index 8b8b7a433ac20..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1b76bc7a17f573c0d52c07ae9ff4302662ae61f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1b94e19d762ddc33cc4e94c6675d93cbde21e3d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1b94e19d762ddc33cc4e94c6675d93cbde21e3d.hip
deleted file mode 100644
index be3611b8829b5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1b94e19d762ddc33cc4e94c6675d93cbde21e3d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1f40c3421b9ad8cf43940530ec50bcf620058f2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1f40c3421b9ad8cf43940530ec50bcf620058f2.hip
deleted file mode 100644
index 3dd173a4a4950..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1f40c3421b9ad8cf43940530ec50bcf620058f2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1f721a330b2d0fac13b22061616d7b10c0f91e9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1f721a330b2d0fac13b22061616d7b10c0f91e9.hip
deleted file mode 100644
index 1091b95947f51..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c1f721a330b2d0fac13b22061616d7b10c0f91e9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c250ea59ab6e1ee39cce15cbd3f181047cdee31a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c250ea59ab6e1ee39cce15cbd3f181047cdee31a.hip
deleted file mode 100644
index 8e8f17dae6afd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c250ea59ab6e1ee39cce15cbd3f181047cdee31a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2541b6b5cf27de3f45f60671d36602f07ce1783.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2541b6b5cf27de3f45f60671d36602f07ce1783.hip
deleted file mode 100644
index 42638c0f3ad69..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2541b6b5cf27de3f45f60671d36602f07ce1783.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c27b3026f1dc3056dee3a3e64bf31c45683607c9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c27b3026f1dc3056dee3a3e64bf31c45683607c9.hip
deleted file mode 100644
index 913d62130eb99..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c27b3026f1dc3056dee3a3e64bf31c45683607c9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c28de8f96c8315877031a2d56261e95fee6aef44.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c28de8f96c8315877031a2d56261e95fee6aef44.hip
deleted file mode 100644
index c57a9de06d866..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c28de8f96c8315877031a2d56261e95fee6aef44.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c29110dd501853e87ebc122dd1971b0bb1bcd92f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c29110dd501853e87ebc122dd1971b0bb1bcd92f.hip
deleted file mode 100644
index 87a12c06c1d6c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c29110dd501853e87ebc122dd1971b0bb1bcd92f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2940fd05efd52bdf8a3f9aa4b78bde9b5809b34.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2940fd05efd52bdf8a3f9aa4b78bde9b5809b34.hip
deleted file mode 100644
index 80bfad3347f64..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2940fd05efd52bdf8a3f9aa4b78bde9b5809b34.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2a2856bf9a81544a30d535a13554e3a8107c476.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2a2856bf9a81544a30d535a13554e3a8107c476.hip
deleted file mode 100644
index 456dd8659aa32..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2a2856bf9a81544a30d535a13554e3a8107c476.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2b719893a4d8a1e71857966d399f06c0a41749c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2b719893a4d8a1e71857966d399f06c0a41749c.hip
deleted file mode 100644
index 236caac1bde5e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2b719893a4d8a1e71857966d399f06c0a41749c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2f04447e6a94c94a2315454e71d7d607a9fd0f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2f04447e6a94c94a2315454e71d7d607a9fd0f8.hip
deleted file mode 100644
index f141d2d50ecf0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2f04447e6a94c94a2315454e71d7d607a9fd0f8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2fcced07cc194a8050bc7b2f791453b3f5b2064.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2fcced07cc194a8050bc7b2f791453b3f5b2064.hip
deleted file mode 100644
index 1782df95df2e2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c2fcced07cc194a8050bc7b2f791453b3f5b2064.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c323a4d1f24d59bddd20ed2f2fb6446627b0ae8b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c323a4d1f24d59bddd20ed2f2fb6446627b0ae8b.hip
deleted file mode 100644
index 49da9084941d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c323a4d1f24d59bddd20ed2f2fb6446627b0ae8b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c355189ade9b1a8269230232db754a3881b53168.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c355189ade9b1a8269230232db754a3881b53168.hip
deleted file mode 100644
index 7445c3af504c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c355189ade9b1a8269230232db754a3881b53168.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c35ea54eb6cd0f3756c462c66d9be956279b46ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c35ea54eb6cd0f3756c462c66d9be956279b46ad.hip
deleted file mode 100644
index ce40ce2f6da87..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c35ea54eb6cd0f3756c462c66d9be956279b46ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c363ee1b087f6b504a3dd3972b96e77db02b0582.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c363ee1b087f6b504a3dd3972b96e77db02b0582.hip
deleted file mode 100644
index f83da199a20cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c363ee1b087f6b504a3dd3972b96e77db02b0582.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c3cfaf0d53869c373f6d0ec821b008dbb819141a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c3cfaf0d53869c373f6d0ec821b008dbb819141a.hip
deleted file mode 100644
index 5cb2f2c3aac32..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c3cfaf0d53869c373f6d0ec821b008dbb819141a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c3d0eaf9399c863d672e8c08d123739bab837d4b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c3d0eaf9399c863d672e8c08d123739bab837d4b.hip
deleted file mode 100644
index f8badbef3afeb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c3d0eaf9399c863d672e8c08d123739bab837d4b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4015f0d0a7a5173810f6f17c00065e03fc61a89.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4015f0d0a7a5173810f6f17c00065e03fc61a89.hip
deleted file mode 100644
index 6295804a8f976..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4015f0d0a7a5173810f6f17c00065e03fc61a89.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c402e84359b2037a29efd1d6ce7213ba7605ab25.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c402e84359b2037a29efd1d6ce7213ba7605ab25.hip
deleted file mode 100644
index f2617a099b88c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c402e84359b2037a29efd1d6ce7213ba7605ab25.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c41b6eda4f250da059fe0c428428219ff5a250ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c41b6eda4f250da059fe0c428428219ff5a250ef.hip
deleted file mode 100644
index 16787715cd590..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c41b6eda4f250da059fe0c428428219ff5a250ef.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c42ab428503e8f8bfa78c8cb8d9afad9f5185118.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c42ab428503e8f8bfa78c8cb8d9afad9f5185118.hip
deleted file mode 100644
index 7b27936874ded..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c42ab428503e8f8bfa78c8cb8d9afad9f5185118.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4376ac8d82db1bc25fa273a80dfbf8b71ee5e2b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4376ac8d82db1bc25fa273a80dfbf8b71ee5e2b.hip
deleted file mode 100644
index 022a42e53b369..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4376ac8d82db1bc25fa273a80dfbf8b71ee5e2b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c45a5e40f6a66bc5292a56e0097c69fe37cedfb3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c45a5e40f6a66bc5292a56e0097c69fe37cedfb3.hip
deleted file mode 100644
index 91d6fbd9347f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c45a5e40f6a66bc5292a56e0097c69fe37cedfb3.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c487a1a9933239270f44b1e08e1cf5323521c089.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c487a1a9933239270f44b1e08e1cf5323521c089.hip
deleted file mode 100644
index bacc30d483ef7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c487a1a9933239270f44b1e08e1cf5323521c089.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4997f79435cf64add10506acb97d0647cfbb3d4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4997f79435cf64add10506acb97d0647cfbb3d4.hip
deleted file mode 100644
index 0a6649fa1e742..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4997f79435cf64add10506acb97d0647cfbb3d4.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4b34d3cb673447773f6da23e9cf52b98e99f718.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4b34d3cb673447773f6da23e9cf52b98e99f718.hip
deleted file mode 100644
index 8d51e201f81d8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4b34d3cb673447773f6da23e9cf52b98e99f718.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4c3425fe683d35dc3335db77d183ad1620b7a92.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4c3425fe683d35dc3335db77d183ad1620b7a92.hip
deleted file mode 100644
index 98c92d20d8a0b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4c3425fe683d35dc3335db77d183ad1620b7a92.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4c6c405cefe204824e8fad1b3dd34bba87e796a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4c6c405cefe204824e8fad1b3dd34bba87e796a.hip
deleted file mode 100644
index ee753848c646c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4c6c405cefe204824e8fad1b3dd34bba87e796a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4de1bc135191f3c2aff740f4c6bb7e98da42f84.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4de1bc135191f3c2aff740f4c6bb7e98da42f84.hip
deleted file mode 100644
index 9d8e124a5c5f1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4de1bc135191f3c2aff740f4c6bb7e98da42f84.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4dec99707511cebd9188d216ee0a148d729b470.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4dec99707511cebd9188d216ee0a148d729b470.hip
deleted file mode 100644
index cfc5294375bef..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c4dec99707511cebd9188d216ee0a148d729b470.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c538dc4f65d02776875627cbd20a9c794d70b043.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c538dc4f65d02776875627cbd20a9c794d70b043.hip
deleted file mode 100644
index ea23fa40a6f48..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c538dc4f65d02776875627cbd20a9c794d70b043.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c53e295b68e807774ed31bb914e4bc59312a77d7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c53e295b68e807774ed31bb914e4bc59312a77d7.hip
deleted file mode 100644
index 0f2e34ea0bdb7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c53e295b68e807774ed31bb914e4bc59312a77d7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c56aa150611b0d4800470c1493dc907082a5c23f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c56aa150611b0d4800470c1493dc907082a5c23f.hip
deleted file mode 100644
index a9e3d97b44a80..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c56aa150611b0d4800470c1493dc907082a5c23f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c581974c8b6f43f60d0af29c350d850b55c03121.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c581974c8b6f43f60d0af29c350d850b55c03121.hip
deleted file mode 100644
index 529e8114dc5ab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c581974c8b6f43f60d0af29c350d850b55c03121.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59937be2b9a13d6520fdcc922e4e75c9fa085ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59937be2b9a13d6520fdcc922e4e75c9fa085ab.hip
deleted file mode 100644
index ae94ab9da8d2b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59937be2b9a13d6520fdcc922e4e75c9fa085ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59a22c6efd8bb8815887325aa0b739e260cc754.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59a22c6efd8bb8815887325aa0b739e260cc754.hip
deleted file mode 100644
index 17e79e74eec39..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59a22c6efd8bb8815887325aa0b739e260cc754.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59ab718fa23f24f09a713ac28a339208a7a5802.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59ab718fa23f24f09a713ac28a339208a7a5802.hip
deleted file mode 100644
index 9056885194941..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c59ab718fa23f24f09a713ac28a339208a7a5802.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5b440ca9a5196ee1e72c878c87d96934e9273c8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5b440ca9a5196ee1e72c878c87d96934e9273c8.hip
deleted file mode 100644
index 3fc61faf14bad..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5b440ca9a5196ee1e72c878c87d96934e9273c8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5fcdea177734366d3bf283317a65cc3fffda611.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5fcdea177734366d3bf283317a65cc3fffda611.hip
deleted file mode 100644
index 513bb8eab599e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5fcdea177734366d3bf283317a65cc3fffda611.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5fef330a975002ed15670e8e7b26a10376d3cb7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5fef330a975002ed15670e8e7b26a10376d3cb7.hip
deleted file mode 100644
index 5ae228165bd69..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c5fef330a975002ed15670e8e7b26a10376d3cb7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c64f4cdce32189065362a502105c31bd2d9d99a4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c64f4cdce32189065362a502105c31bd2d9d99a4.hip
deleted file mode 100644
index 40fde5dd421f5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c64f4cdce32189065362a502105c31bd2d9d99a4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c6e2da8b791d31f4ba05ef5f833fd6dea9e35f1c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c6e2da8b791d31f4ba05ef5f833fd6dea9e35f1c.hip
deleted file mode 100644
index 131ecc94e5382..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c6e2da8b791d31f4ba05ef5f833fd6dea9e35f1c.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c7568e11e44ce70924d27e683190422cfae5c31d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c7568e11e44ce70924d27e683190422cfae5c31d.hip
deleted file mode 100644
index d09808107a789..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c7568e11e44ce70924d27e683190422cfae5c31d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c7af2bbfac25de2853be344b9f636226c1c0112d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c7af2bbfac25de2853be344b9f636226c1c0112d.hip
deleted file mode 100644
index f938c131727b0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c7af2bbfac25de2853be344b9f636226c1c0112d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c806d7803d06ef8aac1d5caac9f36aafd47653d5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c806d7803d06ef8aac1d5caac9f36aafd47653d5.hip
deleted file mode 100644
index 94661ebbba9ae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c806d7803d06ef8aac1d5caac9f36aafd47653d5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c80dce1a17d073259250ec0c87ade69e639ffa8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c80dce1a17d073259250ec0c87ade69e639ffa8e.hip
deleted file mode 100644
index cc0ee6ec65302..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c80dce1a17d073259250ec0c87ade69e639ffa8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c8dbfaffc8a9b573f194f9c63f1175d9725f8950.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c8dbfaffc8a9b573f194f9c63f1175d9725f8950.hip
deleted file mode 100644
index ca5e85e996fde..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c8dbfaffc8a9b573f194f9c63f1175d9725f8950.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c8f6461673882d636772ae4d26e78eabcb568f31.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c8f6461673882d636772ae4d26e78eabcb568f31.hip
deleted file mode 100644
index a2d5ba81161fc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c8f6461673882d636772ae4d26e78eabcb568f31.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c919b8ed877d4244d01a17ecb948b459e361ff24.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c919b8ed877d4244d01a17ecb948b459e361ff24.hip
deleted file mode 100644
index f0f9c3733cf6b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c919b8ed877d4244d01a17ecb948b459e361ff24.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c921a4790f982d48bcaf950123c699647afb739b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c921a4790f982d48bcaf950123c699647afb739b.hip
deleted file mode 100644
index dc436e44e37e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c921a4790f982d48bcaf950123c699647afb739b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9312d7159369d13f3148a6f0882dfad6921ceec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9312d7159369d13f3148a6f0882dfad6921ceec.hip
deleted file mode 100644
index b3256b0f5c6a1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9312d7159369d13f3148a6f0882dfad6921ceec.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9530e20038eb40c49bc8b045be0cf4e7e6b4eac.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9530e20038eb40c49bc8b045be0cf4e7e6b4eac.hip
deleted file mode 100644
index 2e96f767ae26a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9530e20038eb40c49bc8b045be0cf4e7e6b4eac.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c977735a36c325706bd19a12df66ed0839b032b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c977735a36c325706bd19a12df66ed0839b032b1.hip
deleted file mode 100644
index 6d56b66518f3a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c977735a36c325706bd19a12df66ed0839b032b1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9ad71883a19b522486706d3705700c012a6fc19.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9ad71883a19b522486706d3705700c012a6fc19.hip
deleted file mode 100644
index 411be43aef395..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9ad71883a19b522486706d3705700c012a6fc19.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9ba0a3369d4e4eaea1c902a90e6501f232dd57c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9ba0a3369d4e4eaea1c902a90e6501f232dd57c.hip
deleted file mode 100644
index 49a9b123990cf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9ba0a3369d4e4eaea1c902a90e6501f232dd57c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9f1e7e478a2208c4d32e2d7e6abebdc16bcc5fe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9f1e7e478a2208c4d32e2d7e6abebdc16bcc5fe.hip
deleted file mode 100644
index cdd3aa0ddde32..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9f1e7e478a2208c4d32e2d7e6abebdc16bcc5fe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9f28230817c9d9805c41dfcd4e834fe302e1df1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9f28230817c9d9805c41dfcd4e834fe302e1df1.hip
deleted file mode 100644
index 0c350a844d3b0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9f28230817c9d9805c41dfcd4e834fe302e1df1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9fb8343e623e46f01893a2b61345d1ca5928671.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9fb8343e623e46f01893a2b61345d1ca5928671.hip
deleted file mode 100644
index 5f4eb47dab4df..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9fb8343e623e46f01893a2b61345d1ca5928671.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9fe51f982abd60e567d4238d3266fb60e45814b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9fe51f982abd60e567d4238d3266fb60e45814b.hip
deleted file mode 100644
index 012cc9795d4a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_c9fe51f982abd60e567d4238d3266fb60e45814b.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca00cfdc5592b7440d72482a18781e9cf3afb05a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca00cfdc5592b7440d72482a18781e9cf3afb05a.hip
deleted file mode 100644
index 759d7de093d6d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca00cfdc5592b7440d72482a18781e9cf3afb05a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca1992a2634cd6674076611be54197c715ad8271.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca1992a2634cd6674076611be54197c715ad8271.hip
deleted file mode 100644
index 81942aff4b819..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca1992a2634cd6674076611be54197c715ad8271.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca3975efd767ddf7c12e308d948bdcaf0968493a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca3975efd767ddf7c12e308d948bdcaf0968493a.hip
deleted file mode 100644
index eb4ddb914fc30..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca3975efd767ddf7c12e308d948bdcaf0968493a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca3d98ff43fbb80ceb82fc22ab039bee898969b0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca3d98ff43fbb80ceb82fc22ab039bee898969b0.hip
deleted file mode 100644
index 3a0c6d636142d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca3d98ff43fbb80ceb82fc22ab039bee898969b0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca4c6ad28aff1976c6dd36974ec3b339aa3090e9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca4c6ad28aff1976c6dd36974ec3b339aa3090e9.hip
deleted file mode 100644
index f3d182ec80611..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca4c6ad28aff1976c6dd36974ec3b339aa3090e9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca5681d4e5871aacef74bdba9e368445875252d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca5681d4e5871aacef74bdba9e368445875252d3.hip
deleted file mode 100644
index 39af1d4f8e71b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca5681d4e5871aacef74bdba9e368445875252d3.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca920c3239bb5796b1ab2fc75177eb3b820aa784.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca920c3239bb5796b1ab2fc75177eb3b820aa784.hip
deleted file mode 100644
index fb307c260c41a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ca920c3239bb5796b1ab2fc75177eb3b820aa784.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cabb7b12cdd9b8b522af577e13232b2459dbd38d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cabb7b12cdd9b8b522af577e13232b2459dbd38d.hip
deleted file mode 100644
index bfbacdddbd4e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cabb7b12cdd9b8b522af577e13232b2459dbd38d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cae6c7efbfc831e2bcfc8c1efa1a486c02627cbf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cae6c7efbfc831e2bcfc8c1efa1a486c02627cbf.hip
deleted file mode 100644
index 4d85ee9b645e6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cae6c7efbfc831e2bcfc8c1efa1a486c02627cbf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_caede7a18f3e3d5e24f6c70392413a2cda16ac15.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_caede7a18f3e3d5e24f6c70392413a2cda16ac15.hip
deleted file mode 100644
index 06f1e87193b5e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_caede7a18f3e3d5e24f6c70392413a2cda16ac15.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb10303a0b79f2710eb7c66896d3c1f8b12c04dd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb10303a0b79f2710eb7c66896d3c1f8b12c04dd.hip
deleted file mode 100644
index 69efe0f023a60..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb10303a0b79f2710eb7c66896d3c1f8b12c04dd.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1a0ce432c27f4cfa51731c3ef181bf60c8a727.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1a0ce432c27f4cfa51731c3ef181bf60c8a727.hip
deleted file mode 100644
index b6809394551f7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1a0ce432c27f4cfa51731c3ef181bf60c8a727.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1b91c16e0255fe7a0a85638b98d94634e143a9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1b91c16e0255fe7a0a85638b98d94634e143a9.hip
deleted file mode 100644
index 36ded3caa1604..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1b91c16e0255fe7a0a85638b98d94634e143a9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1deea4f4fab0db31d46a91228601f0c272d6e6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1deea4f4fab0db31d46a91228601f0c272d6e6.hip
deleted file mode 100644
index b5a8f46578590..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb1deea4f4fab0db31d46a91228601f0c272d6e6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb20538073888bdb3174a8e9c32d7449072aa753.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb20538073888bdb3174a8e9c32d7449072aa753.hip
deleted file mode 100644
index 9034069a0ee81..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb20538073888bdb3174a8e9c32d7449072aa753.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb3d5273945c5d40cc05c2660af2df1fb7a15f3c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb3d5273945c5d40cc05c2660af2df1fb7a15f3c.hip
deleted file mode 100644
index 704f387c313c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb3d5273945c5d40cc05c2660af2df1fb7a15f3c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb4576e8ea5d59d7663f3760009a00a19e1b0667.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb4576e8ea5d59d7663f3760009a00a19e1b0667.hip
deleted file mode 100644
index 8ece9799f4254..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cb4576e8ea5d59d7663f3760009a00a19e1b0667.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbd571f4fe576fdb17d5f75a558cb6747087c7f2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbd571f4fe576fdb17d5f75a558cb6747087c7f2.hip
deleted file mode 100644
index 1ce00dac81640..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbd571f4fe576fdb17d5f75a558cb6747087c7f2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbe5a98163e878c7697e554758ebd0597c2c1760.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbe5a98163e878c7697e554758ebd0597c2c1760.hip
deleted file mode 100644
index 3b92dead1a98a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbe5a98163e878c7697e554758ebd0597c2c1760.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbf3e4d4d4837a0cb33b78c4f2767b1d93da0850.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbf3e4d4d4837a0cb33b78c4f2767b1d93da0850.hip
deleted file mode 100644
index 464b0174d1ddc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cbf3e4d4d4837a0cb33b78c4f2767b1d93da0850.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc127a63d56099e08125b16939dac82f0173122b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc127a63d56099e08125b16939dac82f0173122b.hip
deleted file mode 100644
index 78939c36b2e1a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc127a63d56099e08125b16939dac82f0173122b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc4ac5a18f57f2ebb65f7e356e858ab0d59b2133.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc4ac5a18f57f2ebb65f7e356e858ab0d59b2133.hip
deleted file mode 100644
index c2c61945ec381..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc4ac5a18f57f2ebb65f7e356e858ab0d59b2133.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc54b107e1b557ea36b5cbaf7fe3dfce05415c86.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc54b107e1b557ea36b5cbaf7fe3dfce05415c86.hip
deleted file mode 100644
index 7555947b0ba67..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cc54b107e1b557ea36b5cbaf7fe3dfce05415c86.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ccac6c0e61b65c9422c7f30fbd979031698370a9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ccac6c0e61b65c9422c7f30fbd979031698370a9.hip
deleted file mode 100644
index 34ae1f2682ca1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ccac6c0e61b65c9422c7f30fbd979031698370a9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ccd0b777df1328bf24e070ed4cdf8615bb2199fe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ccd0b777df1328bf24e070ed4cdf8615bb2199fe.hip
deleted file mode 100644
index e134e6ea51682..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ccd0b777df1328bf24e070ed4cdf8615bb2199fe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd0453a5c3828c1358360f31f5d3b7258e17fdb9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd0453a5c3828c1358360f31f5d3b7258e17fdb9.hip
deleted file mode 100644
index 435630ecd9dc4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd0453a5c3828c1358360f31f5d3b7258e17fdb9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd4efcdd12184211c74e7b3f2f30fecf1041ca32.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd4efcdd12184211c74e7b3f2f30fecf1041ca32.hip
deleted file mode 100644
index 75a9e30603d2e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd4efcdd12184211c74e7b3f2f30fecf1041ca32.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd757a8bbeabd16a44d149ab188430f6d79ddcaf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd757a8bbeabd16a44d149ab188430f6d79ddcaf.hip
deleted file mode 100644
index 1933aed734763..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cd757a8bbeabd16a44d149ab188430f6d79ddcaf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cde0582e1aef74f9209de638b553ec0671476258.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cde0582e1aef74f9209de638b553ec0671476258.hip
deleted file mode 100644
index 198c46f1f6c7d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cde0582e1aef74f9209de638b553ec0671476258.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce4714e4f33340859c106a3129993e22652262e2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce4714e4f33340859c106a3129993e22652262e2.hip
deleted file mode 100644
index 4de81137449b0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce4714e4f33340859c106a3129993e22652262e2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5064e27ba427cb951f7e1b01328b0beb6b2b7c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5064e27ba427cb951f7e1b01328b0beb6b2b7c.hip
deleted file mode 100644
index 667194c237421..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5064e27ba427cb951f7e1b01328b0beb6b2b7c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5ad502dd40353312d561e9f40aa478c16ef5b1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5ad502dd40353312d561e9f40aa478c16ef5b1.hip
deleted file mode 100644
index 7289ddad2559f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5ad502dd40353312d561e9f40aa478c16ef5b1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5b5932f6df9a194ceb0d69220fba9596528eec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5b5932f6df9a194ceb0d69220fba9596528eec.hip
deleted file mode 100644
index 7b0e36bfc9dee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5b5932f6df9a194ceb0d69220fba9596528eec.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5c161b725becf059fb4439c668edd454ac77d1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5c161b725becf059fb4439c668edd454ac77d1.hip
deleted file mode 100644
index c595a779cdea8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce5c161b725becf059fb4439c668edd454ac77d1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce909cb5f96a4884caa0d2eb8c5e6bc7fa352797.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce909cb5f96a4884caa0d2eb8c5e6bc7fa352797.hip
deleted file mode 100644
index 16d62af317432..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ce909cb5f96a4884caa0d2eb8c5e6bc7fa352797.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ceb9544e2a0caae2c9e3dd8bbd2c509e8dca1379.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ceb9544e2a0caae2c9e3dd8bbd2c509e8dca1379.hip
deleted file mode 100644
index 2e46e916aac97..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ceb9544e2a0caae2c9e3dd8bbd2c509e8dca1379.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cee81ab2e2678816c7b516d2d4c50e8cb5874c68.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cee81ab2e2678816c7b516d2d4c50e8cb5874c68.hip
deleted file mode 100644
index b587145a630d4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cee81ab2e2678816c7b516d2d4c50e8cb5874c68.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cf5c6c0bfaf98f6e655fc443246b81fcc730fe97.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cf5c6c0bfaf98f6e655fc443246b81fcc730fe97.hip
deleted file mode 100644
index 306d9691df38d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cf5c6c0bfaf98f6e655fc443246b81fcc730fe97.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cf73e1fc0015094861ca0c1c81bacdbe0c5b8f37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cf73e1fc0015094861ca0c1c81bacdbe0c5b8f37.hip
deleted file mode 100644
index ccbb0403a4cdf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cf73e1fc0015094861ca0c1c81bacdbe0c5b8f37.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cfda56a4eb08b803332f25bda6209932d9624acc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cfda56a4eb08b803332f25bda6209932d9624acc.hip
deleted file mode 100644
index 9d8de36c23888..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cfda56a4eb08b803332f25bda6209932d9624acc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cfec97bdfb6fa95e057eaf5a8138853e1c0884f2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cfec97bdfb6fa95e057eaf5a8138853e1c0884f2.hip
deleted file mode 100644
index 4975d2a08b390..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_cfec97bdfb6fa95e057eaf5a8138853e1c0884f2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d00f65bc99ca08eba66564d34f72f2769bff9491.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d00f65bc99ca08eba66564d34f72f2769bff9491.hip
deleted file mode 100644
index 153be1294071a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d00f65bc99ca08eba66564d34f72f2769bff9491.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d036096f49a89730f8af7e75457c88cb8ae64165.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d036096f49a89730f8af7e75457c88cb8ae64165.hip
deleted file mode 100644
index f9e1d3a26fa10..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d036096f49a89730f8af7e75457c88cb8ae64165.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d049a1b8f4c1c6d37973ce38593efda1de8ce0cd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d049a1b8f4c1c6d37973ce38593efda1de8ce0cd.hip
deleted file mode 100644
index 0333b3fbd6d6c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d049a1b8f4c1c6d37973ce38593efda1de8ce0cd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d04dc4ed02eb42c3fe303342801ed3073a0dcb8e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d04dc4ed02eb42c3fe303342801ed3073a0dcb8e.hip
deleted file mode 100644
index f4bbd80c1a8c7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d04dc4ed02eb42c3fe303342801ed3073a0dcb8e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d06ba4c996570ddab77b6ff1e2a0101b638543eb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d06ba4c996570ddab77b6ff1e2a0101b638543eb.hip
deleted file mode 100644
index c1041812f4602..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d06ba4c996570ddab77b6ff1e2a0101b638543eb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0863830fc5d43dc6d6400280e892bb7de2892d4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0863830fc5d43dc6d6400280e892bb7de2892d4.hip
deleted file mode 100644
index 980649f7afd2d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0863830fc5d43dc6d6400280e892bb7de2892d4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d090b771a4f9750132f549c82a88b4ab00dce5c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d090b771a4f9750132f549c82a88b4ab00dce5c7.hip
deleted file mode 100644
index 6af910f2d430e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d090b771a4f9750132f549c82a88b4ab00dce5c7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0b09e8513646fbb2a007544a63ec9e2b04dc4c2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0b09e8513646fbb2a007544a63ec9e2b04dc4c2.hip
deleted file mode 100644
index 09a6a7583fd97..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0b09e8513646fbb2a007544a63ec9e2b04dc4c2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0daa59f5dce6fc3965193ae37d8c82a3d1834e6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0daa59f5dce6fc3965193ae37d8c82a3d1834e6.hip
deleted file mode 100644
index aa1a8da5be924..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0daa59f5dce6fc3965193ae37d8c82a3d1834e6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0dd0165ee91c095a19ceddf08789e3576912590.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0dd0165ee91c095a19ceddf08789e3576912590.hip
deleted file mode 100644
index 01a5d42cc233e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0dd0165ee91c095a19ceddf08789e3576912590.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0de618ff3ea9f67b90f2227fb7fcc74ea34183d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0de618ff3ea9f67b90f2227fb7fcc74ea34183d.hip
deleted file mode 100644
index 976954f04e904..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0de618ff3ea9f67b90f2227fb7fcc74ea34183d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0f63cafbeb445408c884727b473667fb479675e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0f63cafbeb445408c884727b473667fb479675e.hip
deleted file mode 100644
index e8c968a70cae5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d0f63cafbeb445408c884727b473667fb479675e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d137b7b6e04e1caf43a62bd6788a75361cfa98f6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d137b7b6e04e1caf43a62bd6788a75361cfa98f6.hip
deleted file mode 100644
index b8a62d8016523..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d137b7b6e04e1caf43a62bd6788a75361cfa98f6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1840494c4fa78ff399c0399b3ad7ca3d22d4587.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1840494c4fa78ff399c0399b3ad7ca3d22d4587.hip
deleted file mode 100644
index 424d1718b8dba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1840494c4fa78ff399c0399b3ad7ca3d22d4587.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d18727988e47264b42b4153dc82fc1a750f08db0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d18727988e47264b42b4153dc82fc1a750f08db0.hip
deleted file mode 100644
index 8bda0c0c3e3de..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d18727988e47264b42b4153dc82fc1a750f08db0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1c0dfd19a08d61586758091370acbdc6f267017.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1c0dfd19a08d61586758091370acbdc6f267017.hip
deleted file mode 100644
index af5115b8125e4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1c0dfd19a08d61586758091370acbdc6f267017.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1c25cfc437d8bd803860e39a45b2f3b9fa48393.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1c25cfc437d8bd803860e39a45b2f3b9fa48393.hip
deleted file mode 100644
index c9eb1a48c620b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1c25cfc437d8bd803860e39a45b2f3b9fa48393.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1d3eacc320104100bce46235fe656e5a8223c66.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1d3eacc320104100bce46235fe656e5a8223c66.hip
deleted file mode 100644
index 7fd064e617556..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d1d3eacc320104100bce46235fe656e5a8223c66.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d20d45aa85c0daa299da98c277cee826fe67bd27.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d20d45aa85c0daa299da98c277cee826fe67bd27.hip
deleted file mode 100644
index 6c7e14a3772c5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d20d45aa85c0daa299da98c277cee826fe67bd27.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d257148f457557ea80ca56690e525db3a4b0ff55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d257148f457557ea80ca56690e525db3a4b0ff55.hip
deleted file mode 100644
index 4b9b9aa6d8acf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d257148f457557ea80ca56690e525db3a4b0ff55.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d25ce4b3e9cc392ceafebc7fe3bcbe05aaad4bbc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d25ce4b3e9cc392ceafebc7fe3bcbe05aaad4bbc.hip
deleted file mode 100644
index 517bebd1cc9ac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d25ce4b3e9cc392ceafebc7fe3bcbe05aaad4bbc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2d08c5470a385d0160b2c1441fd1c30fff1c17c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2d08c5470a385d0160b2c1441fd1c30fff1c17c.hip
deleted file mode 100644
index 9fff84645693e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2d08c5470a385d0160b2c1441fd1c30fff1c17c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2daccc4b3a0f90bff39cb4597f8b7e484613d9e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2daccc4b3a0f90bff39cb4597f8b7e484613d9e.hip
deleted file mode 100644
index c5f1fd582aa86..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2daccc4b3a0f90bff39cb4597f8b7e484613d9e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2dfdb42c1b380e860aa5609302f29698dd27923.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2dfdb42c1b380e860aa5609302f29698dd27923.hip
deleted file mode 100644
index 94ed9dd287989..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2dfdb42c1b380e860aa5609302f29698dd27923.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2f4b869ff23874b6bde0aab68c419108b7e69f4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2f4b869ff23874b6bde0aab68c419108b7e69f4.hip
deleted file mode 100644
index d3fdd90b305e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d2f4b869ff23874b6bde0aab68c419108b7e69f4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d32c64ef01aa228277d031a74df51363f98aa2b0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d32c64ef01aa228277d031a74df51363f98aa2b0.hip
deleted file mode 100644
index a0051666d9ccf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d32c64ef01aa228277d031a74df51363f98aa2b0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d34d6cdcd81a456125ab5e0875466c6334d8e5c8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d34d6cdcd81a456125ab5e0875466c6334d8e5c8.hip
deleted file mode 100644
index e0ae2c571360b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d34d6cdcd81a456125ab5e0875466c6334d8e5c8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d34fcb56caa8f80404789fba0ffac447483a4d84.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d34fcb56caa8f80404789fba0ffac447483a4d84.hip
deleted file mode 100644
index bcbfb7d4ece14..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d34fcb56caa8f80404789fba0ffac447483a4d84.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3784fb4c0685d7b651f4113f3c71e050881f3a5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3784fb4c0685d7b651f4113f3c71e050881f3a5.hip
deleted file mode 100644
index 5b5162d095f96..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3784fb4c0685d7b651f4113f3c71e050881f3a5.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3a23ded424200d0c6f06b1dbd0a7b7b0e7b5d9b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3a23ded424200d0c6f06b1dbd0a7b7b0e7b5d9b.hip
deleted file mode 100644
index 1c6be7bb35a22..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3a23ded424200d0c6f06b1dbd0a7b7b0e7b5d9b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3a2edf232786d458e2125f8dfeda8847f842afa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3a2edf232786d458e2125f8dfeda8847f842afa.hip
deleted file mode 100644
index f5c3c75ac36bb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3a2edf232786d458e2125f8dfeda8847f842afa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3af8763f289dace1054bdcb4dfeda28b0aefcae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3af8763f289dace1054bdcb4dfeda28b0aefcae.hip
deleted file mode 100644
index e17ac519f030f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3af8763f289dace1054bdcb4dfeda28b0aefcae.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3fce1e11aee2273620e75efe4aa0390fcde9ba5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3fce1e11aee2273620e75efe4aa0390fcde9ba5.hip
deleted file mode 100644
index 09ef59db3945f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d3fce1e11aee2273620e75efe4aa0390fcde9ba5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d40569ae9dbd693c0ab3d6ba69704d31e451011b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d40569ae9dbd693c0ab3d6ba69704d31e451011b.hip
deleted file mode 100644
index f38468cb1d8d6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d40569ae9dbd693c0ab3d6ba69704d31e451011b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d41b6a64dd181f2efa65aaed03a3d229b3566c1d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d41b6a64dd181f2efa65aaed03a3d229b3566c1d.hip
deleted file mode 100644
index fcccaaa3538fd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d41b6a64dd181f2efa65aaed03a3d229b3566c1d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d41cd6b60a97e7071518cbd1a63abb8b910df024.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d41cd6b60a97e7071518cbd1a63abb8b910df024.hip
deleted file mode 100644
index 32007f72daa1f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d41cd6b60a97e7071518cbd1a63abb8b910df024.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d43715cce8935439f90172d141050d78c7e76fb7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d43715cce8935439f90172d141050d78c7e76fb7.hip
deleted file mode 100644
index ab3595779079f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d43715cce8935439f90172d141050d78c7e76fb7.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4605b2ad3e3753c5f255678abc1690b949c5abc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4605b2ad3e3753c5f255678abc1690b949c5abc.hip
deleted file mode 100644
index bf74552890ad0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4605b2ad3e3753c5f255678abc1690b949c5abc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4645b713821371161a9925dec8a3d6c157ba1aa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4645b713821371161a9925dec8a3d6c157ba1aa.hip
deleted file mode 100644
index 3bdb63e15dcbd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4645b713821371161a9925dec8a3d6c157ba1aa.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4aff499ad527be5fe33b8e92547df57af26d40d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4aff499ad527be5fe33b8e92547df57af26d40d.hip
deleted file mode 100644
index 235361844edbc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4aff499ad527be5fe33b8e92547df57af26d40d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4b99af9a573df50a27fccbec3fa8e350f1854eb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4b99af9a573df50a27fccbec3fa8e350f1854eb.hip
deleted file mode 100644
index dbb25419bf47b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4b99af9a573df50a27fccbec3fa8e350f1854eb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4c9f975891087e6eed6393629b41155deafc509.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4c9f975891087e6eed6393629b41155deafc509.hip
deleted file mode 100644
index 24b01131cdd28..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d4c9f975891087e6eed6393629b41155deafc509.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d50ac8e8a03f8e7ec2c6e993dd39f09f465dab57.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d50ac8e8a03f8e7ec2c6e993dd39f09f465dab57.hip
deleted file mode 100644
index 99ca50c1d7ff3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d50ac8e8a03f8e7ec2c6e993dd39f09f465dab57.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d54ac01458df3f240e0656d82330f9de23ba9651.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d54ac01458df3f240e0656d82330f9de23ba9651.hip
deleted file mode 100644
index a15c0881098d6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d54ac01458df3f240e0656d82330f9de23ba9651.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d54b3731883a5f8393d60d27487f8d017aedd3f9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d54b3731883a5f8393d60d27487f8d017aedd3f9.hip
deleted file mode 100644
index beae00879d5bb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d54b3731883a5f8393d60d27487f8d017aedd3f9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d5e82799f4452e148c3e02acd6526cf30757eb52.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d5e82799f4452e148c3e02acd6526cf30757eb52.hip
deleted file mode 100644
index 6c2885c16cdd5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d5e82799f4452e148c3e02acd6526cf30757eb52.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d5edfe3e3dc3008b928c8e6dbd50784b905f189e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d5edfe3e3dc3008b928c8e6dbd50784b905f189e.hip
deleted file mode 100644
index c9b9deb0b994d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d5edfe3e3dc3008b928c8e6dbd50784b905f189e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d600779c17b7b21c18e1308e6d765fe02a7945d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d600779c17b7b21c18e1308e6d765fe02a7945d3.hip
deleted file mode 100644
index dd4cce764ed3e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d600779c17b7b21c18e1308e6d765fe02a7945d3.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d6149eea92f2c40c11de3b778102fcf9b6a006b8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d6149eea92f2c40c11de3b778102fcf9b6a006b8.hip
deleted file mode 100644
index bb7dd91588903..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d6149eea92f2c40c11de3b778102fcf9b6a006b8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d623b36cc3f56d1001b2d3abadd8a5628fefd014.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d623b36cc3f56d1001b2d3abadd8a5628fefd014.hip
deleted file mode 100644
index ba05dc9911f1c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d623b36cc3f56d1001b2d3abadd8a5628fefd014.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d63c8c746055851217a514321cd735eaf6937263.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d63c8c746055851217a514321cd735eaf6937263.hip
deleted file mode 100644
index 8d06bb2fa1874..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d63c8c746055851217a514321cd735eaf6937263.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d64b8b52f4a98801e185e2f132b2f80c29dd0c37.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d64b8b52f4a98801e185e2f132b2f80c29dd0c37.hip
deleted file mode 100644
index 9d3145849aa52..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d64b8b52f4a98801e185e2f132b2f80c29dd0c37.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d66b79c4ebdcfd239cecec58203606bc123bd6bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d66b79c4ebdcfd239cecec58203606bc123bd6bb.hip
deleted file mode 100644
index 820d61d7280d7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d66b79c4ebdcfd239cecec58203606bc123bd6bb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d66c30148a6fa816937f2f095802264d3dfa0273.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d66c30148a6fa816937f2f095802264d3dfa0273.hip
deleted file mode 100644
index 7ab541dd60cff..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d66c30148a6fa816937f2f095802264d3dfa0273.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d703eea8075cacec4d41fee7dc4734f593ee79e8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d703eea8075cacec4d41fee7dc4734f593ee79e8.hip
deleted file mode 100644
index c6c6cf7160998..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d703eea8075cacec4d41fee7dc4734f593ee79e8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d712f23ef88ae5d7b161d36f42d22a5ba53b6354.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d712f23ef88ae5d7b161d36f42d22a5ba53b6354.hip
deleted file mode 100644
index 08547b31058da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d712f23ef88ae5d7b161d36f42d22a5ba53b6354.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d713fe25dc90b3511fc259cebf463376dcb55d84.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d713fe25dc90b3511fc259cebf463376dcb55d84.hip
deleted file mode 100644
index bb768bb887582..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d713fe25dc90b3511fc259cebf463376dcb55d84.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7145383e39dec0e346b5094401acf85ef3c2075.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7145383e39dec0e346b5094401acf85ef3c2075.hip
deleted file mode 100644
index c323c090710b2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7145383e39dec0e346b5094401acf85ef3c2075.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d723b191785c97d284675f700a7baeb52a2eb791.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d723b191785c97d284675f700a7baeb52a2eb791.hip
deleted file mode 100644
index 219a5fc831356..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d723b191785c97d284675f700a7baeb52a2eb791.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7290cc4c3036c9205e689cbcc60e7d16b97a7d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7290cc4c3036c9205e689cbcc60e7d16b97a7d6.hip
deleted file mode 100644
index 07ac1999a0c2b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7290cc4c3036c9205e689cbcc60e7d16b97a7d6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d733f4c03e338ea7c6d8f759c1132499bdcea059.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d733f4c03e338ea7c6d8f759c1132499bdcea059.hip
deleted file mode 100644
index 97536b5743c0c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d733f4c03e338ea7c6d8f759c1132499bdcea059.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d773df9ccfc1ace90fe3afb5c00976deabedf6f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d773df9ccfc1ace90fe3afb5c00976deabedf6f8.hip
deleted file mode 100644
index 14bff89c4bdea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d773df9ccfc1ace90fe3afb5c00976deabedf6f8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7adde8780b39f1364c572a19c3bfb19417678e3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7adde8780b39f1364c572a19c3bfb19417678e3.hip
deleted file mode 100644
index 8e20942df4586..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7adde8780b39f1364c572a19c3bfb19417678e3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7bda8157fb27d544e049fd7d2ec735725f1bf44.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7bda8157fb27d544e049fd7d2ec735725f1bf44.hip
deleted file mode 100644
index 6d62663289417..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7bda8157fb27d544e049fd7d2ec735725f1bf44.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7fae2c18645d36a181a0bdd2d8ca7a4ac0f6d1d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7fae2c18645d36a181a0bdd2d8ca7a4ac0f6d1d.hip
deleted file mode 100644
index 1b9cd96ceb9f6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d7fae2c18645d36a181a0bdd2d8ca7a4ac0f6d1d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d82773721479613ad72e334510a248f1436b38d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d82773721479613ad72e334510a248f1436b38d6.hip
deleted file mode 100644
index 7ec972e200808..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d82773721479613ad72e334510a248f1436b38d6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d867098db97b3f26e71a151c63b74260bfab21f8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d867098db97b3f26e71a151c63b74260bfab21f8.hip
deleted file mode 100644
index fdc9ff3d9fd69..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d867098db97b3f26e71a151c63b74260bfab21f8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d86e4dcbe9c4cac8f7c8c5d97ce384ae0cbdbfbc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d86e4dcbe9c4cac8f7c8c5d97ce384ae0cbdbfbc.hip
deleted file mode 100644
index 8e100e7b34ab0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d86e4dcbe9c4cac8f7c8c5d97ce384ae0cbdbfbc.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d8901a63986cc28ef24cab012b32114851a8c1ec.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d8901a63986cc28ef24cab012b32114851a8c1ec.hip
deleted file mode 100644
index d3deccfe241a3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d8901a63986cc28ef24cab012b32114851a8c1ec.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9061c204d8a85c974676f4438994a0be9d69a60.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9061c204d8a85c974676f4438994a0be9d69a60.hip
deleted file mode 100644
index 58f4dc0004ba5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9061c204d8a85c974676f4438994a0be9d69a60.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d924ee32b178b6bffa7a71603d6e2818f66177a5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d924ee32b178b6bffa7a71603d6e2818f66177a5.hip
deleted file mode 100644
index bf00a987c4d38..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d924ee32b178b6bffa7a71603d6e2818f66177a5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d937609afa8e21a761dad6b01ff3f26346e450fc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d937609afa8e21a761dad6b01ff3f26346e450fc.hip
deleted file mode 100644
index a4d16e3cdd9c1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d937609afa8e21a761dad6b01ff3f26346e450fc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d95835bc6f000d3a3379bbc38d90e83dcaf867ee.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d95835bc6f000d3a3379bbc38d90e83dcaf867ee.hip
deleted file mode 100644
index 5eb2be47469ab..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d95835bc6f000d3a3379bbc38d90e83dcaf867ee.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d992eab7de49033f5480c5e86a69e675db0d2a19.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d992eab7de49033f5480c5e86a69e675db0d2a19.hip
deleted file mode 100644
index 0a8788862dd73..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d992eab7de49033f5480c5e86a69e675db0d2a19.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9c23b7f8fcc4e4f4c81f5f00cfd345b98df2e0f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9c23b7f8fcc4e4f4c81f5f00cfd345b98df2e0f.hip
deleted file mode 100644
index 26cd69469ea22..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9c23b7f8fcc4e4f4c81f5f00cfd345b98df2e0f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9c3e27b522320dcca5ee84fa534b03aae2bfea9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9c3e27b522320dcca5ee84fa534b03aae2bfea9.hip
deleted file mode 100644
index 27e2cec912a6e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_d9c3e27b522320dcca5ee84fa534b03aae2bfea9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da07d8b5666423da30a95e3b2cabd3839d200981.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da07d8b5666423da30a95e3b2cabd3839d200981.hip
deleted file mode 100644
index 882eb06599611..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da07d8b5666423da30a95e3b2cabd3839d200981.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da29a515d14dac02066bcd4701285b9916b43cf5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da29a515d14dac02066bcd4701285b9916b43cf5.hip
deleted file mode 100644
index 868b3bcfef645..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da29a515d14dac02066bcd4701285b9916b43cf5.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da6afccdee4107507a64323e17bf12c46da2b92a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da6afccdee4107507a64323e17bf12c46da2b92a.hip
deleted file mode 100644
index f2a99dce82043..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da6afccdee4107507a64323e17bf12c46da2b92a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da74887afedbd67928fe4d596709f9ff92530611.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da74887afedbd67928fe4d596709f9ff92530611.hip
deleted file mode 100644
index bd3dd21914c57..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da74887afedbd67928fe4d596709f9ff92530611.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da822ea727fb3543e445e4000f7e6ebb946d6a3b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da822ea727fb3543e445e4000f7e6ebb946d6a3b.hip
deleted file mode 100644
index d73043aa58d93..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da822ea727fb3543e445e4000f7e6ebb946d6a3b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da9f6e1d59132fe96709490af25bd794f267851c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da9f6e1d59132fe96709490af25bd794f267851c.hip
deleted file mode 100644
index e75388caf2d4d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_da9f6e1d59132fe96709490af25bd794f267851c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db0d0cf55d90b3f3c9eecada1db93c420f34b1ae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db0d0cf55d90b3f3c9eecada1db93c420f34b1ae.hip
deleted file mode 100644
index 69c56122a64b4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db0d0cf55d90b3f3c9eecada1db93c420f34b1ae.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db5016bff9e5dc37184d2b9417eb351c7ea1c322.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db5016bff9e5dc37184d2b9417eb351c7ea1c322.hip
deleted file mode 100644
index 8f34d4d0813da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db5016bff9e5dc37184d2b9417eb351c7ea1c322.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db85839ee8d464c5a81b8dad9839f5e0f4b467a8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db85839ee8d464c5a81b8dad9839f5e0f4b467a8.hip
deleted file mode 100644
index fbb400e9cfc79..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db85839ee8d464c5a81b8dad9839f5e0f4b467a8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db8f0bd93b352d28c5b6d78f4332026993f0bea4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db8f0bd93b352d28c5b6d78f4332026993f0bea4.hip
deleted file mode 100644
index 96ecc14b53963..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_db8f0bd93b352d28c5b6d78f4332026993f0bea4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbae1670fac6812b2d2cbad973e4b475509ea504.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbae1670fac6812b2d2cbad973e4b475509ea504.hip
deleted file mode 100644
index 8315731de9398..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbae1670fac6812b2d2cbad973e4b475509ea504.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbb06b43d5d65429e23cc717448cf1fffb0cfd74.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbb06b43d5d65429e23cc717448cf1fffb0cfd74.hip
deleted file mode 100644
index e3328b73adb49..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbb06b43d5d65429e23cc717448cf1fffb0cfd74.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbc4135fce01e8731fec7a78d0cc0fdeeae28b90.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbc4135fce01e8731fec7a78d0cc0fdeeae28b90.hip
deleted file mode 100644
index a95d754746133..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbc4135fce01e8731fec7a78d0cc0fdeeae28b90.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbcea8f7b5930abf76eecefce92d0db785d2df5d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbcea8f7b5930abf76eecefce92d0db785d2df5d.hip
deleted file mode 100644
index 9bd274029d5dc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbcea8f7b5930abf76eecefce92d0db785d2df5d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbde2ef18e2174ebe13a6e7c8c2a6b05a6612047.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbde2ef18e2174ebe13a6e7c8c2a6b05a6612047.hip
deleted file mode 100644
index 06855f1bd70c9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dbde2ef18e2174ebe13a6e7c8c2a6b05a6612047.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc039d422a57c159ea4dbcc867d766ff1b356a07.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc039d422a57c159ea4dbcc867d766ff1b356a07.hip
deleted file mode 100644
index 9c93a89103663..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc039d422a57c159ea4dbcc867d766ff1b356a07.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc08afbff5def8bcb4e823657ce01f57c9dc77c9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc08afbff5def8bcb4e823657ce01f57c9dc77c9.hip
deleted file mode 100644
index 949824c472e54..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc08afbff5def8bcb4e823657ce01f57c9dc77c9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc184767d723f4995791848cdc68bd948408204f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc184767d723f4995791848cdc68bd948408204f.hip
deleted file mode 100644
index 0c1b6d1f19eb9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc184767d723f4995791848cdc68bd948408204f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc1a7f9b1afeba6690fdc0d0d1755ea89c805573.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc1a7f9b1afeba6690fdc0d0d1755ea89c805573.hip
deleted file mode 100644
index c8a9ebbeb0867..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc1a7f9b1afeba6690fdc0d0d1755ea89c805573.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc34b6ef496d4e0d8fbbe10731d4a7b1c136c036.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc34b6ef496d4e0d8fbbe10731d4a7b1c136c036.hip
deleted file mode 100644
index cfc31f9c7da1d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc34b6ef496d4e0d8fbbe10731d4a7b1c136c036.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc3d625c5ad3e871f5a727ac946df642d988b9ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc3d625c5ad3e871f5a727ac946df642d988b9ab.hip
deleted file mode 100644
index 4d1fd757a5d7a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc3d625c5ad3e871f5a727ac946df642d988b9ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc4d27535b9570b8f4b790470a83c1d0a9a2b6ce.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc4d27535b9570b8f4b790470a83c1d0a9a2b6ce.hip
deleted file mode 100644
index af903c96af62c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc4d27535b9570b8f4b790470a83c1d0a9a2b6ce.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc5ba6d73f331c76e696953606c5b347b6a46f3f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc5ba6d73f331c76e696953606c5b347b6a46f3f.hip
deleted file mode 100644
index 80eb63e6db380..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc5ba6d73f331c76e696953606c5b347b6a46f3f.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc62a8db637d32e7dfdb2521cbdae6e1fbbd5fd1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc62a8db637d32e7dfdb2521cbdae6e1fbbd5fd1.hip
deleted file mode 100644
index 043b84f9f3005..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc62a8db637d32e7dfdb2521cbdae6e1fbbd5fd1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc818f3ce244743cb1dbff9aca399df90742a6d0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc818f3ce244743cb1dbff9aca399df90742a6d0.hip
deleted file mode 100644
index 2b28b28a6e946..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc818f3ce244743cb1dbff9aca399df90742a6d0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc91797c1474a368e9cb056b50b4629d7736c3cb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc91797c1474a368e9cb056b50b4629d7736c3cb.hip
deleted file mode 100644
index 2a222d6728c23..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc91797c1474a368e9cb056b50b4629d7736c3cb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc9e54273c0ea2358fb573a7d918aa7b09fe07f9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc9e54273c0ea2358fb573a7d918aa7b09fe07f9.hip
deleted file mode 100644
index 265006b2cb949..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dc9e54273c0ea2358fb573a7d918aa7b09fe07f9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dcf815ef540060cc7ed43e1c57a28e1d080c5621.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dcf815ef540060cc7ed43e1c57a28e1d080c5621.hip
deleted file mode 100644
index 4a27c80049d55..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dcf815ef540060cc7ed43e1c57a28e1d080c5621.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd10bbf37503bbc92af82bc3487989b41b20ca85.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd10bbf37503bbc92af82bc3487989b41b20ca85.hip
deleted file mode 100644
index 82b2ecd178cd7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd10bbf37503bbc92af82bc3487989b41b20ca85.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd11806cd2d3ef1127f676b2d98bf8fff2a1e5ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd11806cd2d3ef1127f676b2d98bf8fff2a1e5ab.hip
deleted file mode 100644
index 7187346120013..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd11806cd2d3ef1127f676b2d98bf8fff2a1e5ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd35634440edb25cb095800b882c70aaceca1dbb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd35634440edb25cb095800b882c70aaceca1dbb.hip
deleted file mode 100644
index e4810cf301271..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd35634440edb25cb095800b882c70aaceca1dbb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd67d442001d2b167e70e8730abde4d4461b8569.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd67d442001d2b167e70e8730abde4d4461b8569.hip
deleted file mode 100644
index 94216509787d4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd67d442001d2b167e70e8730abde4d4461b8569.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd9494d9ac35eba6794a4f9120d2db9932596ef8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd9494d9ac35eba6794a4f9120d2db9932596ef8.hip
deleted file mode 100644
index 2fb0d4ba5f50a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dd9494d9ac35eba6794a4f9120d2db9932596ef8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, true,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dda8d021381083bc48b7fb1840729254dd8e5137.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dda8d021381083bc48b7fb1840729254dd8e5137.hip
deleted file mode 100644
index 2f36a58ffe77b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dda8d021381083bc48b7fb1840729254dd8e5137.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ddcb1cfea1b0dbe50a02252cba99428fd977527e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ddcb1cfea1b0dbe50a02252cba99428fd977527e.hip
deleted file mode 100644
index 1ae834a694259..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ddcb1cfea1b0dbe50a02252cba99428fd977527e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dde93ffe7fca311e136e42fbcd12b05c9fc7174c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dde93ffe7fca311e136e42fbcd12b05c9fc7174c.hip
deleted file mode 100644
index b68ef694340bf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dde93ffe7fca311e136e42fbcd12b05c9fc7174c.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ddf5339054f47d9ed6cc7f9e66ab21ce3bccf3db.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ddf5339054f47d9ed6cc7f9e66ab21ce3bccf3db.hip
deleted file mode 100644
index 0e0c167c3b608..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ddf5339054f47d9ed6cc7f9e66ab21ce3bccf3db.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de1ff66d2aeb47d2fdccaa4bb6b9d066b380c99e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de1ff66d2aeb47d2fdccaa4bb6b9d066b380c99e.hip
deleted file mode 100644
index 0926fc5eef321..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de1ff66d2aeb47d2fdccaa4bb6b9d066b380c99e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de26a187c4db06115072a5132e1166b5b03368b0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de26a187c4db06115072a5132e1166b5b03368b0.hip
deleted file mode 100644
index 2bb11c53165bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de26a187c4db06115072a5132e1166b5b03368b0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de36bc309877917a18fd21acb30563c7e2f233c1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de36bc309877917a18fd21acb30563c7e2f233c1.hip
deleted file mode 100644
index 0d63a90fde926..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de36bc309877917a18fd21acb30563c7e2f233c1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de5359f0fba3da9dfed06ddbea8fe2a33a9cf40c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de5359f0fba3da9dfed06ddbea8fe2a33a9cf40c.hip
deleted file mode 100644
index e27f787455d1b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de5359f0fba3da9dfed06ddbea8fe2a33a9cf40c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de6683d175affaa5ff261ab8503f64172d8eba8b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de6683d175affaa5ff261ab8503f64172d8eba8b.hip
deleted file mode 100644
index 1c3096ece2388..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de6683d175affaa5ff261ab8503f64172d8eba8b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de7eb562a7eff31d589e12945d80233aac202ae2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de7eb562a7eff31d589e12945d80233aac202ae2.hip
deleted file mode 100644
index 34c1fc87b8357..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de7eb562a7eff31d589e12945d80233aac202ae2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de85901d66dc04b1143bb6404445baf65693b781.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de85901d66dc04b1143bb6404445baf65693b781.hip
deleted file mode 100644
index 9cd6fbd8d4136..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_de85901d66dc04b1143bb6404445baf65693b781.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_deb9ec2cccab94920e40f62a1f0f094acd919d07.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_deb9ec2cccab94920e40f62a1f0f094acd919d07.hip
deleted file mode 100644
index f7f8414dc77a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_deb9ec2cccab94920e40f62a1f0f094acd919d07.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df0b2bcba57e77d975ec5304fc50cbd09cddf4bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df0b2bcba57e77d975ec5304fc50cbd09cddf4bb.hip
deleted file mode 100644
index 8c0ab4ea85f11..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df0b2bcba57e77d975ec5304fc50cbd09cddf4bb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df4bb75ca79f805a81fbad750ad22f6d22b0d8ff.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df4bb75ca79f805a81fbad750ad22f6d22b0d8ff.hip
deleted file mode 100644
index 2558cbe0b9e25..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df4bb75ca79f805a81fbad750ad22f6d22b0d8ff.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df4c9eb48da49a61957537270d94e56cb4e426be.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df4c9eb48da49a61957537270d94e56cb4e426be.hip
deleted file mode 100644
index 6c8837fbec7da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df4c9eb48da49a61957537270d94e56cb4e426be.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df5b1c6758d4b8540158299dd0362297083084c2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df5b1c6758d4b8540158299dd0362297083084c2.hip
deleted file mode 100644
index 7231dd59a6ad3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df5b1c6758d4b8540158299dd0362297083084c2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df645b3888dc8d1df50c47c0d75822eebd3eb019.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df645b3888dc8d1df50c47c0d75822eebd3eb019.hip
deleted file mode 100644
index 8d96d0d04aacf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df645b3888dc8d1df50c47c0d75822eebd3eb019.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df66feebc9a0dcc508ce002c255154622875e524.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df66feebc9a0dcc508ce002c255154622875e524.hip
deleted file mode 100644
index e8fb2a96b9b55..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_df66feebc9a0dcc508ce002c255154622875e524.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dfcd68acfca68d1acac94f493e25be0ef20f209f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dfcd68acfca68d1acac94f493e25be0ef20f209f.hip
deleted file mode 100644
index e3c865be35f8a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_dfcd68acfca68d1acac94f493e25be0ef20f209f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e02a198f23c409b715761b702d7b0e6e5992701f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e02a198f23c409b715761b702d7b0e6e5992701f.hip
deleted file mode 100644
index 11b33022ae576..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e02a198f23c409b715761b702d7b0e6e5992701f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e035773419a9b3631698a3d375d829af55f7731e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e035773419a9b3631698a3d375d829af55f7731e.hip
deleted file mode 100644
index 26061b55ae0a2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e035773419a9b3631698a3d375d829af55f7731e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e088f0f7363804cf5403adef70828ab32d09a02a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e088f0f7363804cf5403adef70828ab32d09a02a.hip
deleted file mode 100644
index e78f750aeb77b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e088f0f7363804cf5403adef70828ab32d09a02a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e0966fa1ff013e477b1706928de6cb7f8587c154.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e0966fa1ff013e477b1706928de6cb7f8587c154.hip
deleted file mode 100644
index c09d32372ad4c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e0966fa1ff013e477b1706928de6cb7f8587c154.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e09d9baa269dfbb30b714389d1733be51cc419b7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e09d9baa269dfbb30b714389d1733be51cc419b7.hip
deleted file mode 100644
index d2e2e921b0659..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e09d9baa269dfbb30b714389d1733be51cc419b7.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        64,
-        256,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<256,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e0e48d7edfe9513f24ad9fae68cac3aa940b17dd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e0e48d7edfe9513f24ad9fae68cac3aa940b17dd.hip
deleted file mode 100644
index 8b37aef49222a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e0e48d7edfe9513f24ad9fae68cac3aa940b17dd.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e10f47a44400de385ddbeb99475b717c5646fb41.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e10f47a44400de385ddbeb99475b717c5646fb41.hip
deleted file mode 100644
index 0ce7a98dc0c89..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e10f47a44400de385ddbeb99475b717c5646fb41.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e11a3b7d4fdfed64e64f7a95dbc64eff541092d6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e11a3b7d4fdfed64e64f7a95dbc64eff541092d6.hip
deleted file mode 100644
index 5ec322eafb110..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e11a3b7d4fdfed64e64f7a95dbc64eff541092d6.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e13b86fe4e153e0bfa8d1e75f3641fe32b0c5149.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e13b86fe4e153e0bfa8d1e75f3641fe32b0c5149.hip
deleted file mode 100644
index 19a8ab7b04d8b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e13b86fe4e153e0bfa8d1e75f3641fe32b0c5149.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e16075c3a5fcfe63ba12e854bb1fed6873f014ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e16075c3a5fcfe63ba12e854bb1fed6873f014ab.hip
deleted file mode 100644
index 1330304ff7148..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e16075c3a5fcfe63ba12e854bb1fed6873f014ab.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e16edb824cecf459a8ec51b8dc74b1e06369aceb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e16edb824cecf459a8ec51b8dc74b1e06369aceb.hip
deleted file mode 100644
index 98120391c7081..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e16edb824cecf459a8ec51b8dc74b1e06369aceb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1c1a31a1d8556cbe0b6ea76faacc78855108539.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1c1a31a1d8556cbe0b6ea76faacc78855108539.hip
deleted file mode 100644
index 3f2d3b26b68e3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1c1a31a1d8556cbe0b6ea76faacc78855108539.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1cc934ba7baab1a2eb062df1e4ee5066e9ffbc3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1cc934ba7baab1a2eb062df1e4ee5066e9ffbc3.hip
deleted file mode 100644
index d7b52329688ed..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1cc934ba7baab1a2eb062df1e4ee5066e9ffbc3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1d85ad2c9d197f501267fe0804e6985802fbd18.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1d85ad2c9d197f501267fe0804e6985802fbd18.hip
deleted file mode 100644
index 716331a828805..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e1d85ad2c9d197f501267fe0804e6985802fbd18.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2762543d3380185e304f84749a70db1b8d3dd8c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2762543d3380185e304f84749a70db1b8d3dd8c.hip
deleted file mode 100644
index afd102f264fc0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2762543d3380185e304f84749a70db1b8d3dd8c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e28fd64c2f2b27577109a984e6ab82f5f0fcb296.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e28fd64c2f2b27577109a984e6ab82f5f0fcb296.hip
deleted file mode 100644
index cfeae1a8441f7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e28fd64c2f2b27577109a984e6ab82f5f0fcb296.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2b629c37cf94134693ce455b8c88b72a39df7fe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2b629c37cf94134693ce455b8c88b72a39df7fe.hip
deleted file mode 100644
index 711f8949ebb26..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2b629c37cf94134693ce455b8c88b72a39df7fe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2bf6805a489739abb77c13173d57723e9304afa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2bf6805a489739abb77c13173d57723e9304afa.hip
deleted file mode 100644
index 909734e02eaa2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2bf6805a489739abb77c13173d57723e9304afa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2c9f955f227430c6224ebc347649386be7f01eb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2c9f955f227430c6224ebc347649386be7f01eb.hip
deleted file mode 100644
index fc0ed99b6c01f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2c9f955f227430c6224ebc347649386be7f01eb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2deafd2f36cee29109fb824e0135407453adcfe.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2deafd2f36cee29109fb824e0135407453adcfe.hip
deleted file mode 100644
index 73401b787b4d3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e2deafd2f36cee29109fb824e0135407453adcfe.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e3015c5d50481547aa5754d042d9d7040cf1c7ff.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e3015c5d50481547aa5754d042d9d7040cf1c7ff.hip
deleted file mode 100644
index 6455762cb9475..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e3015c5d50481547aa5754d042d9d7040cf1c7ff.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e307a1b0d5a8f94e0a0f4032f401d20b4b643523.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e307a1b0d5a8f94e0a0f4032f401d20b4b643523.hip
deleted file mode 100644
index 3d43ac4cf01fb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e307a1b0d5a8f94e0a0f4032f401d20b4b643523.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e334e691714f0b99773c2ac515ed82de0f387065.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e334e691714f0b99773c2ac515ed82de0f387065.hip
deleted file mode 100644
index 3839ba35429d7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e334e691714f0b99773c2ac515ed82de0f387065.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e34b7e452a4db74189334697e3a240ad68085f0e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e34b7e452a4db74189334697e3a240ad68085f0e.hip
deleted file mode 100644
index b6268ae2e48e1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e34b7e452a4db74189334697e3a240ad68085f0e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e389d0e4442cd8304081892ddc75043e68a6398c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e389d0e4442cd8304081892ddc75043e68a6398c.hip
deleted file mode 100644
index 9e34eaf3da342..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e389d0e4442cd8304081892ddc75043e68a6398c.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e465193d97d43237c22c04478ca5833011d8dc8b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e465193d97d43237c22c04478ca5833011d8dc8b.hip
deleted file mode 100644
index e0219c5cd464c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e465193d97d43237c22c04478ca5833011d8dc8b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e477abef05ff37ec27705eda51896e2aa3a04966.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e477abef05ff37ec27705eda51896e2aa3a04966.hip
deleted file mode 100644
index 2adad72235169..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e477abef05ff37ec27705eda51896e2aa3a04966.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e4d9a2396ceccdadab24602f30e9070901a76dc7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e4d9a2396ceccdadab24602f30e9070901a76dc7.hip
deleted file mode 100644
index efd272520172b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e4d9a2396ceccdadab24602f30e9070901a76dc7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e502730dea6987e2c038446c448aa08bdcc23113.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e502730dea6987e2c038446c448aa08bdcc23113.hip
deleted file mode 100644
index 57155cc00fccd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e502730dea6987e2c038446c448aa08bdcc23113.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e514c6b4bc75d95a150104a17972abae77cb47ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e514c6b4bc75d95a150104a17972abae77cb47ed.hip
deleted file mode 100644
index 6a5b2c6d13c97..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e514c6b4bc75d95a150104a17972abae77cb47ed.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e52e3053f30f780f346fa6b7a836ad2554cb85df.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e52e3053f30f780f346fa6b7a836ad2554cb85df.hip
deleted file mode 100644
index d1402916ab42e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e52e3053f30f780f346fa6b7a836ad2554cb85df.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e56757fb17f5e94a6ba1fb14540a68c36d571159.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e56757fb17f5e94a6ba1fb14540a68c36d571159.hip
deleted file mode 100644
index 2635cbc893667..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e56757fb17f5e94a6ba1fb14540a68c36d571159.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e578ec9e09d3b78dca6b5bf0be1538657f02f319.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e578ec9e09d3b78dca6b5bf0be1538657f02f319.hip
deleted file mode 100644
index 35c85a1f9d39b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e578ec9e09d3b78dca6b5bf0be1538657f02f319.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5935fbda313d3518f142f43d46f56c600f69286.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5935fbda313d3518f142f43d46f56c600f69286.hip
deleted file mode 100644
index 9c8ff3e0f0eac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5935fbda313d3518f142f43d46f56c600f69286.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5b2bb9f8466de1ad5210e4c39ee7b8ecacdffa9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5b2bb9f8466de1ad5210e4c39ee7b8ecacdffa9.hip
deleted file mode 100644
index c3f716b6782d6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5b2bb9f8466de1ad5210e4c39ee7b8ecacdffa9.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5b65fc519ea7cfcd19f7eddbc3acad6842ff558.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5b65fc519ea7cfcd19f7eddbc3acad6842ff558.hip
deleted file mode 100644
index 9287b1cfd1791..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5b65fc519ea7cfcd19f7eddbc3acad6842ff558.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5c5079636a4a31a849ce8a5af89d50330a74628.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5c5079636a4a31a849ce8a5af89d50330a74628.hip
deleted file mode 100644
index 04d0a50da47d9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5c5079636a4a31a849ce8a5af89d50330a74628.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5ccd5f7ddc894b2717112cbfc766804e02b7bd1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5ccd5f7ddc894b2717112cbfc766804e02b7bd1.hip
deleted file mode 100644
index d98d14f90e26e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e5ccd5f7ddc894b2717112cbfc766804e02b7bd1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e618fb4e529104fc90069c8779ce5463460bd516.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e618fb4e529104fc90069c8779ce5463460bd516.hip
deleted file mode 100644
index 6c66a8aba8af4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e618fb4e529104fc90069c8779ce5463460bd516.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e638053e01268a4c5883620fc6a9901951e2e01a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e638053e01268a4c5883620fc6a9901951e2e01a.hip
deleted file mode 100644
index d8b3b85504706..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e638053e01268a4c5883620fc6a9901951e2e01a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e639a1e84faa98477b05df71d363b9ff0f9b2760.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e639a1e84faa98477b05df71d363b9ff0f9b2760.hip
deleted file mode 100644
index c85bc59841b06..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e639a1e84faa98477b05df71d363b9ff0f9b2760.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e68a9e05debd456a9975953f7b0d510e7a0f6978.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e68a9e05debd456a9975953f7b0d510e7a0f6978.hip
deleted file mode 100644
index 7e48c597d5e63..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e68a9e05debd456a9975953f7b0d510e7a0f6978.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6973d75297bd2c3432a7c88e8a9ee1c9ae693bf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6973d75297bd2c3432a7c88e8a9ee1c9ae693bf.hip
deleted file mode 100644
index 3a2d6375b87f5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6973d75297bd2c3432a7c88e8a9ee1c9ae693bf.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6b53fb8d81148ff384d31a703bb4c2e7a5a33af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6b53fb8d81148ff384d31a703bb4c2e7a5a33af.hip
deleted file mode 100644
index 18cf40ec5aa12..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6b53fb8d81148ff384d31a703bb4c2e7a5a33af.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6e0ec1db1ea308e226f675e68e29b839e41b252.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6e0ec1db1ea308e226f675e68e29b839e41b252.hip
deleted file mode 100644
index 0d7760b463a68..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6e0ec1db1ea308e226f675e68e29b839e41b252.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6e6b10e73733716e71ebf5a53703fb935fc5e02.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6e6b10e73733716e71ebf5a53703fb935fc5e02.hip
deleted file mode 100644
index 8acda06dd9019..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e6e6b10e73733716e71ebf5a53703fb935fc5e02.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7153f9a9b0b7c54ddf2debbe297efcffbb4fcfa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7153f9a9b0b7c54ddf2debbe297efcffbb4fcfa.hip
deleted file mode 100644
index 507e4ca2ff1f4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7153f9a9b0b7c54ddf2debbe297efcffbb4fcfa.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    128,
-    true,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<128, ck_tile::bf16_t, true, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e73a776ae4ba68c23acab1a5a6381684051738ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e73a776ae4ba68c23acab1a5a6381684051738ab.hip
deleted file mode 100644
index d0c25a7283f61..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e73a776ae4ba68c23acab1a5a6381684051738ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e75c757c67aa23cb88e1aced6fcf36b7b28391db.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e75c757c67aa23cb88e1aced6fcf36b7b28391db.hip
deleted file mode 100644
index 1e8b6398d9e2d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e75c757c67aa23cb88e1aced6fcf36b7b28391db.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e75d492ac3a6ab75648056bcf26250a4aa929cfd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e75d492ac3a6ab75648056bcf26250a4aa929cfd.hip
deleted file mode 100644
index b9b534ef39a4c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e75d492ac3a6ab75648056bcf26250a4aa929cfd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e76879f8ff4796f48ad87ff8003f4f6e6adca9a0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e76879f8ff4796f48ad87ff8003f4f6e6adca9a0.hip
deleted file mode 100644
index f5253a7b60c10..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e76879f8ff4796f48ad87ff8003f4f6e6adca9a0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7ae1294b6dea5c8b93c2b814fa7460c4047105b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7ae1294b6dea5c8b93c2b814fa7460c4047105b.hip
deleted file mode 100644
index 982c1f7ca71c6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7ae1294b6dea5c8b93c2b814fa7460c4047105b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7b2eb64b66d46359fab44333c2c484f4c9dd5de.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7b2eb64b66d46359fab44333c2c484f4c9dd5de.hip
deleted file mode 100644
index aa64d12f513bc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7b2eb64b66d46359fab44333c2c484f4c9dd5de.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7c0a99e949baa5f3a7ee2d6e84427982f82f76d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7c0a99e949baa5f3a7ee2d6e84427982f82f76d.hip
deleted file mode 100644
index b8bacaa883def..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7c0a99e949baa5f3a7ee2d6e84427982f82f76d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7d37e7ee96c392fa24c02a9143438a3a7d05741.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7d37e7ee96c392fa24c02a9143438a3a7d05741.hip
deleted file mode 100644
index dc213af2ce759..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7d37e7ee96c392fa24c02a9143438a3a7d05741.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7de729aa50c10d8101ef504138c3769e3286753.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7de729aa50c10d8101ef504138c3769e3286753.hip
deleted file mode 100644
index 35e14e0d3c31e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e7de729aa50c10d8101ef504138c3769e3286753.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e83c604d1b8260958becd1c7c209745ff9151715.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e83c604d1b8260958becd1c7c209745ff9151715.hip
deleted file mode 100644
index d78d224e97479..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e83c604d1b8260958becd1c7c209745ff9151715.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e89bcea4393593313d18a4aa6dcb44cd75bc828d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e89bcea4393593313d18a4aa6dcb44cd75bc828d.hip
deleted file mode 100644
index 5e069502bc59c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e89bcea4393593313d18a4aa6dcb44cd75bc828d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8a9427f34bbf5ddb28a39161acc36806e68f2d0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8a9427f34bbf5ddb28a39161acc36806e68f2d0.hip
deleted file mode 100644
index 5fd258cffbb26..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8a9427f34bbf5ddb28a39161acc36806e68f2d0.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8d8fe5f4f8641998b8b805a20b2ca92d019ee59.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8d8fe5f4f8641998b8b805a20b2ca92d019ee59.hip
deleted file mode 100644
index 403288f4f73fb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8d8fe5f4f8641998b8b805a20b2ca92d019ee59.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8d9b65558398c0c10127b560807578ef117d7ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8d9b65558398c0c10127b560807578ef117d7ed.hip
deleted file mode 100644
index 1827de25165b0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e8d9b65558398c0c10127b560807578ef117d7ed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e907e8d1089557dfcc95a05160be5092e9119a53.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e907e8d1089557dfcc95a05160be5092e9119a53.hip
deleted file mode 100644
index 88fe0d2cf3e42..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e907e8d1089557dfcc95a05160be5092e9119a53.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e95e3908479965856843317c8b0c42a6961dfd23.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e95e3908479965856843317c8b0c42a6961dfd23.hip
deleted file mode 100644
index f53424f68a444..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e95e3908479965856843317c8b0c42a6961dfd23.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e986d5f8d5591f3e0f1cdfad19c38c420fd93023.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e986d5f8d5591f3e0f1cdfad19c38c420fd93023.hip
deleted file mode 100644
index cc361a2b88129..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e986d5f8d5591f3e0f1cdfad19c38c420fd93023.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e9b04e6d5527ba0b8089ba8bdd264e2d5759338b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e9b04e6d5527ba0b8089ba8bdd264e2d5759338b.hip
deleted file mode 100644
index 3cc03123fc08f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e9b04e6d5527ba0b8089ba8bdd264e2d5759338b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e9b53fa68641f45baabf40b7cfb8b35a9a1b9c7f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e9b53fa68641f45baabf40b7cfb8b35a9a1b9c7f.hip
deleted file mode 100644
index b7ff6e3311873..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_e9b53fa68641f45baabf40b7cfb8b35a9a1b9c7f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea077e68dbc1bed2dd20a5f4dd35e0cad6330ee4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea077e68dbc1bed2dd20a5f4dd35e0cad6330ee4.hip
deleted file mode 100644
index bf1d988757841..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea077e68dbc1bed2dd20a5f4dd35e0cad6330ee4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea591185b1c5f521023e250a26f742984255b241.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea591185b1c5f521023e250a26f742984255b241.hip
deleted file mode 100644
index 9bdfda3e22a04..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea591185b1c5f521023e250a26f742984255b241.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea62567e9ea16771d8445464c38f5a2931cb355a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea62567e9ea16771d8445464c38f5a2931cb355a.hip
deleted file mode 100644
index ffe6de2cdf765..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea62567e9ea16771d8445464c38f5a2931cb355a.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea6a6d4cc262ea838dbb83ee747112f95fa297bc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea6a6d4cc262ea838dbb83ee747112f95fa297bc.hip
deleted file mode 100644
index 7a0bcef89fbfa..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ea6a6d4cc262ea838dbb83ee747112f95fa297bc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eab6cdc59bf216f7045f0cf5f221bb91ec415cd2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eab6cdc59bf216f7045f0cf5f221bb91ec415cd2.hip
deleted file mode 100644
index f77d15767190f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eab6cdc59bf216f7045f0cf5f221bb91ec415cd2.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eac353f963c52624cf79e82cc2b2c02eed94b677.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eac353f963c52624cf79e82cc2b2c02eed94b677.hip
deleted file mode 100644
index ae04c653a625c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eac353f963c52624cf79e82cc2b2c02eed94b677.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eac5952f46f4f2bf06257b00661774eeed48a323.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eac5952f46f4f2bf06257b00661774eeed48a323.hip
deleted file mode 100644
index d0b9029b79e15..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eac5952f46f4f2bf06257b00661774eeed48a323.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<false, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, false, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eb278488b2cca114adca5e4614d86f92447f937a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eb278488b2cca114adca5e4614d86f92447f937a.hip
deleted file mode 100644
index 53912552f8cee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eb278488b2cca114adca5e4614d86f92447f937a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ebb241b947a0adfc8e50c5d71765c14af24593ae.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ebb241b947a0adfc8e50c5d71765c14af24593ae.hip
deleted file mode 100644
index ff1b8b1812cb8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ebb241b947a0adfc8e50c5d71765c14af24593ae.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ebb9abf5b09e63cbe76390bb46ff7cbefb3141f0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ebb9abf5b09e63cbe76390bb46ff7cbefb3141f0.hip
deleted file mode 100644
index 3750b4dfaa396..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ebb9abf5b09e63cbe76390bb46ff7cbefb3141f0.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec171210efd217c07d357fcf42e5372ad7e9abab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec171210efd217c07d357fcf42e5372ad7e9abab.hip
deleted file mode 100644
index d1d93065719e7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec171210efd217c07d357fcf42e5372ad7e9abab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec3deb1382003ac010d9bc1c59d1878d3ec7a727.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec3deb1382003ac010d9bc1c59d1878d3ec7a727.hip
deleted file mode 100644
index 7df2d85c8eb64..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec3deb1382003ac010d9bc1c59d1878d3ec7a727.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec51d24ab5f24e003ed6751ae8ae5b327892b15a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec51d24ab5f24e003ed6751ae8ae5b327892b15a.hip
deleted file mode 100644
index d2cfc01a84280..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec51d24ab5f24e003ed6751ae8ae5b327892b15a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec7ec8d547ee9713aa3b5b667f22cdcaa8f62b2d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec7ec8d547ee9713aa3b5b667f22cdcaa8f62b2d.hip
deleted file mode 100644
index 6115f42182896..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec7ec8d547ee9713aa3b5b667f22cdcaa8f62b2d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec7fc24902b1ebd8f2bf8088b0ecf6de8be8362d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec7fc24902b1ebd8f2bf8088b0ecf6de8be8362d.hip
deleted file mode 100644
index 49a98116856a9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec7fc24902b1ebd8f2bf8088b0ecf6de8be8362d.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec9f63a538940e5ace02ae5b5ddc01f730adac4d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec9f63a538940e5ace02ae5b5ddc01f730adac4d.hip
deleted file mode 100644
index e46fb0817921b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ec9f63a538940e5ace02ae5b5ddc01f730adac4d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eca613eaa8471ad7da66d2f8f2b8e07f6e02b467.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eca613eaa8471ad7da66d2f8f2b8e07f6e02b467.hip
deleted file mode 100644
index 160e1e545d2d6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eca613eaa8471ad7da66d2f8f2b8e07f6e02b467.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ecd7dec90b3c62bf3a30bd75d3c6869529a06b01.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ecd7dec90b3c62bf3a30bd75d3c6869529a06b01.hip
deleted file mode 100644
index 14f53fba4a1f9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ecd7dec90b3c62bf3a30bd75d3c6869529a06b01.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ece60111633db08f765b3c7cd5cd768cbd030255.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ece60111633db08f765b3c7cd5cd768cbd030255.hip
deleted file mode 100644
index e3dd992614434..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ece60111633db08f765b3c7cd5cd768cbd030255.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ed37ba962e0288e2840eb0925d016b5a7e3b3164.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ed37ba962e0288e2840eb0925d016b5a7e3b3164.hip
deleted file mode 100644
index 756e8c4ed621d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ed37ba962e0288e2840eb0925d016b5a7e3b3164.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ed6bdf67720e938d538a867548ac3579b8238169.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ed6bdf67720e938d538a867548ac3579b8238169.hip
deleted file mode 100644
index 1b1b714a443ea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ed6bdf67720e938d538a867548ac3579b8238169.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ede81dbc4cb208ef6e684c76ba1eb451d37fe10c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ede81dbc4cb208ef6e684c76ba1eb451d37fe10c.hip
deleted file mode 100644
index 2082ecf146435..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ede81dbc4cb208ef6e684c76ba1eb451d37fe10c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee1a43f2210a8d1e5623411c95c33424cee5e747.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee1a43f2210a8d1e5623411c95c33424cee5e747.hip
deleted file mode 100644
index 21e4b98530be6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee1a43f2210a8d1e5623411c95c33424cee5e747.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee239db5a67c23a383590a651f0d8a0be43a13c7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee239db5a67c23a383590a651f0d8a0be43a13c7.hip
deleted file mode 100644
index ab1e19dfddb79..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee239db5a67c23a383590a651f0d8a0be43a13c7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee8e709eec7aef1fa681053c6d2969a5ff18c45c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee8e709eec7aef1fa681053c6d2969a5ff18c45c.hip
deleted file mode 100644
index 17a2d575fd084..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee8e709eec7aef1fa681053c6d2969a5ff18c45c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee974931e65d6b16b7c868d462b95dcae20b7513.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee974931e65d6b16b7c868d462b95dcae20b7513.hip
deleted file mode 100644
index 5a581e2b944f0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ee974931e65d6b16b7c868d462b95dcae20b7513.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eeb0e96b759e18cf703cfab0cda1385726f6e0a1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eeb0e96b759e18cf703cfab0cda1385726f6e0a1.hip
deleted file mode 100644
index e373e6255556d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eeb0e96b759e18cf703cfab0cda1385726f6e0a1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eee408cf9456ff977aa7d12345e9b2f1e60639f1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eee408cf9456ff977aa7d12345e9b2f1e60639f1.hip
deleted file mode 100644
index ec42f7faf0ec3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_eee408cf9456ff977aa7d12345e9b2f1e60639f1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef2ebb4a86e7ed0001de9c5e607b66fe8877409f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef2ebb4a86e7ed0001de9c5e607b66fe8877409f.hip
deleted file mode 100644
index f4cb467ac3aac..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef2ebb4a86e7ed0001de9c5e607b66fe8877409f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef40f0acf1885096efb840ec5600ec421c4db331.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef40f0acf1885096efb840ec5600ec421c4db331.hip
deleted file mode 100644
index 271b5666abe49..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef40f0acf1885096efb840ec5600ec421c4db331.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           false, false>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::fp16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, false, false, false, false, false, false>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef5421703cbfa63a58ec02701e245d479a1fbfc1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef5421703cbfa63a58ec02701e245d479a1fbfc1.hip
deleted file mode 100644
index 5b62ec8861647..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef5421703cbfa63a58ec02701e245d479a1fbfc1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef7cc2aa1ffd38298b52764a93cd1271b4d92f8d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef7cc2aa1ffd38298b52764a93cd1271b4d92f8d.hip
deleted file mode 100644
index 778bfb8d46089..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ef7cc2aa1ffd38298b52764a93cd1271b4d92f8d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efaa0cb33c71cb8ca7b83dd0e7a6c7b01f6b50a9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efaa0cb33c71cb8ca7b83dd0e7a6c7b01f6b50a9.hip
deleted file mode 100644
index 6c9757987822c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efaa0cb33c71cb8ca7b83dd0e7a6c7b01f6b50a9.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, false, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    256,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<256, ck_tile::bf16_t, false, true, false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efb9e7d9af47cdf79f15f674f8976c05f08b0ce8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efb9e7d9af47cdf79f15f674f8976c05f08b0ce8.hip
deleted file mode 100644
index 3d8e689b434bf..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efb9e7d9af47cdf79f15f674f8976c05f08b0ce8.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efc6a7b25710f0626c3af534111b161e1459d2e1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efc6a7b25710f0626c3af534111b161e1459d2e1.hip
deleted file mode 100644
index eb0e08d17f000..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_efc6a7b25710f0626c3af534111b161e1459d2e1.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f01468c62c878295443981662e037ec5213cf7a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f01468c62c878295443981662e037ec5213cf7a3.hip
deleted file mode 100644
index c1572e5e14af2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f01468c62c878295443981662e037ec5213cf7a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f020134822739be6fa0bb3d98e9dec79f025324a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f020134822739be6fa0bb3d98e9dec79f025324a.hip
deleted file mode 100644
index cd4d0b3d5a93a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f020134822739be6fa0bb3d98e9dec79f025324a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f0209426a8e6bfeef7d8ae7b16db791888142298.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f0209426a8e6bfeef7d8ae7b16db791888142298.hip
deleted file mode 100644
index fd457d4544e94..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f0209426a8e6bfeef7d8ae7b16db791888142298.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f028af9e5e3c25800dde938e991aaab4fc1d64aa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f028af9e5e3c25800dde938e991aaab4fc1d64aa.hip
deleted file mode 100644
index 6be6bb016b280..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f028af9e5e3c25800dde938e991aaab4fc1d64aa.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f053c9c32518b895daaa3521827f37af78836fb8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f053c9c32518b895daaa3521827f37af78836fb8.hip
deleted file mode 100644
index 213bad825ec08..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f053c9c32518b895daaa3521827f37af78836fb8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::fp16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f069b38b26c30bc770f74c856e47eb498f5818e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f069b38b26c30bc770f74c856e47eb498f5818e7.hip
deleted file mode 100644
index 8fc65e43f859a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f069b38b26c30bc770f74c856e47eb498f5818e7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f0cad48d9bc80d58705ea60eb2dda4baad68cedb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f0cad48d9bc80d58705ea60eb2dda4baad68cedb.hip
deleted file mode 100644
index 864b348cc8954..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f0cad48d9bc80d58705ea60eb2dda4baad68cedb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f1246d1013d954a9316f4432c986d3be9459c548.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f1246d1013d954a9316f4432c986d3be9459c548.hip
deleted file mode 100644
index 7dc3a42f6e8b6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f1246d1013d954a9316f4432c986d3be9459c548.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f12f1f1b679cabab04218037ef370d2c7e1fe332.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f12f1f1b679cabab04218037ef370d2c7e1fe332.hip
deleted file mode 100644
index 60c9e624df545..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f12f1f1b679cabab04218037ef370d2c7e1fe332.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f15c41ddb04ec7f80235bb3db19198dd6b699713.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f15c41ddb04ec7f80235bb3db19198dd6b699713.hip
deleted file mode 100644
index e0107ade6e6ba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f15c41ddb04ec7f80235bb3db19198dd6b699713.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f18c74becc24a93427d9c0838784e9b6caad6e81.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f18c74becc24a93427d9c0838784e9b6caad6e81.hip
deleted file mode 100644
index 36b0bc3cce915..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f18c74becc24a93427d9c0838784e9b6caad6e81.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f1ecc90ad7b86791a9e6f73a582aeff30f393804.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f1ecc90ad7b86791a9e6f73a582aeff30f393804.hip
deleted file mode 100644
index 68533e7f9ef9d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f1ecc90ad7b86791a9e6f73a582aeff30f393804.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f21596e8c608a795ff971aea8e199db9e72b65d7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f21596e8c608a795ff971aea8e199db9e72b65d7.hip
deleted file mode 100644
index e40d17a03d406..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f21596e8c608a795ff971aea8e199db9e72b65d7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24bd5b92ce6bba640b8ec6b4e53fe35902c5572.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24bd5b92ce6bba640b8ec6b4e53fe35902c5572.hip
deleted file mode 100644
index bff2fa9414351..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24bd5b92ce6bba640b8ec6b4e53fe35902c5572.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24d42e820adc1a26a428d59df7ffdd7f8580176.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24d42e820adc1a26a428d59df7ffdd7f8580176.hip
deleted file mode 100644
index 2837920826d03..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24d42e820adc1a26a428d59df7ffdd7f8580176.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24f26e45d5cf567d29fbe375fbf8abdec39186f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24f26e45d5cf567d29fbe375fbf8abdec39186f.hip
deleted file mode 100644
index 4de3723f948b2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f24f26e45d5cf567d29fbe375fbf8abdec39186f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f25b87c435bc5d7d85d738f3fdf68947d79f5a77.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f25b87c435bc5d7d85d738f3fdf68947d79f5a77.hip
deleted file mode 100644
index d51c980e8903e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f25b87c435bc5d7d85d738f3fdf68947d79f5a77.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f280e1639680ac1e5830a21f921bfe2cf364ef42.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f280e1639680ac1e5830a21f921bfe2cf364ef42.hip
deleted file mode 100644
index 9aa3ab035db83..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f280e1639680ac1e5830a21f921bfe2cf364ef42.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f2da112b1e07c44fc8a7f19368da203f6935049c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f2da112b1e07c44fc8a7f19368da203f6935049c.hip
deleted file mode 100644
index 927922b0bda9e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f2da112b1e07c44fc8a7f19368da203f6935049c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f30316cfe49323638f71ba688dd8ff9b2266b335.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f30316cfe49323638f71ba688dd8ff9b2266b335.hip
deleted file mode 100644
index 70a90bcb6901b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f30316cfe49323638f71ba688dd8ff9b2266b335.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3193ea266f3718398bc5622f8bc7042c3527a42.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3193ea266f3718398bc5622f8bc7042c3527a42.hip
deleted file mode 100644
index dbbb66f95a2a8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3193ea266f3718398bc5622f8bc7042c3527a42.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f34fdb8294257d951dcc9c4fa7ecf1192568b91b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f34fdb8294257d951dcc9c4fa7ecf1192568b91b.hip
deleted file mode 100644
index e9b21bb104531..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f34fdb8294257d951dcc9c4fa7ecf1192568b91b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f36aaa63ed42a578b953ebd614318d44cf44e8a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f36aaa63ed42a578b953ebd614318d44cf44e8a3.hip
deleted file mode 100644
index 029c41aa8323d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f36aaa63ed42a578b953ebd614318d44cf44e8a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f395bec57c3b2e6e169134dd8d20b287d7405134.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f395bec57c3b2e6e169134dd8d20b287d7405134.hip
deleted file mode 100644
index c7c012e4a802a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f395bec57c3b2e6e169134dd8d20b287d7405134.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3bf7ef503bb026258b3ec3d82d3ef1443046964.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3bf7ef503bb026258b3ec3d82d3ef1443046964.hip
deleted file mode 100644
index 8053a28e2e833..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3bf7ef503bb026258b3ec3d82d3ef1443046964.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3d0166931e4406873d8f552a5d5b61fde2391a3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3d0166931e4406873d8f552a5d5b61fde2391a3.hip
deleted file mode 100644
index 5cb10d9b88def..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3d0166931e4406873d8f552a5d5b61fde2391a3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3fd08d56f8a9be1a8dd104cdb1ac58e283b5064.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3fd08d56f8a9be1a8dd104cdb1ac58e283b5064.hip
deleted file mode 100644
index 1dc20d68b2e3e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3fd08d56f8a9be1a8dd104cdb1ac58e283b5064.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3ff73f82aee3184849d04c2364eaa45c6d0de9c.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3ff73f82aee3184849d04c2364eaa45c6d0de9c.hip
deleted file mode 100644
index d6eff9a6fe4c5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f3ff73f82aee3184849d04c2364eaa45c6d0de9c.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f42cf0e5fe479690883507028748b0cd3dc83cbb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f42cf0e5fe479690883507028748b0cd3dc83cbb.hip
deleted file mode 100644
index ec50bf67d0a80..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f42cf0e5fe479690883507028748b0cd3dc83cbb.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4658c32d562f9d60c5ca1262a2e0df2375063bb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4658c32d562f9d60c5ca1262a2e0df2375063bb.hip
deleted file mode 100644
index 5900808dbbd4d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4658c32d562f9d60c5ca1262a2e0df2375063bb.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f48f8b681a405bfeba5aadaef40f32367ec5cd2b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f48f8b681a405bfeba5aadaef40f32367ec5cd2b.hip
deleted file mode 100644
index db360db1a24f8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f48f8b681a405bfeba5aadaef40f32367ec5cd2b.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<false, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        128,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<128,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             false,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4900c0a5c0d03dc17d7a907ab40652d9920e756.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4900c0a5c0d03dc17d7a907ab40652d9920e756.hip
deleted file mode 100644
index 85ae3594e2852..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4900c0a5c0d03dc17d7a907ab40652d9920e756.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4a6438394dd3427f29aa0bbe58ad1f797c3c38d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4a6438394dd3427f29aa0bbe58ad1f797c3c38d.hip
deleted file mode 100644
index 2e4112941cdfc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4a6438394dd3427f29aa0bbe58ad1f797c3c38d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4b87f983a5e84582efa1663f84da76cf60b5f6f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4b87f983a5e84582efa1663f84da76cf60b5f6f.hip
deleted file mode 100644
index a92c5be29573b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4b87f983a5e84582efa1663f84da76cf60b5f6f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4c803838f5644ccc6f04f7c8a6233fed0b6639e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4c803838f5644ccc6f04f7c8a6233fed0b6639e.hip
deleted file mode 100644
index 223a213ffaac2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4c803838f5644ccc6f04f7c8a6233fed0b6639e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4df1cbfbaf67705820f125b474469ad7ebab0c0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4df1cbfbaf67705820f125b474469ad7ebab0c0.hip
deleted file mode 100644
index 2cc4ee1d2f8d5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f4df1cbfbaf67705820f125b474469ad7ebab0c0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f50fa4ea674a590d0a817367ad9915a5fce20c51.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f50fa4ea674a590d0a817367ad9915a5fce20c51.hip
deleted file mode 100644
index ecddb9f6b4646..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f50fa4ea674a590d0a817367ad9915a5fce20c51.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f51f1a11f778d99a00aa5959a3e58a41fcbfb1e3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f51f1a11f778d99a00aa5959a3e58a41fcbfb1e3.hip
deleted file mode 100644
index eadfc2864ab88..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f51f1a11f778d99a00aa5959a3e58a41fcbfb1e3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f525b59df454ccf53da6cb201e0aa8d09f52a2ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f525b59df454ccf53da6cb201e0aa8d09f52a2ad.hip
deleted file mode 100644
index 6e20a2bd7e1e3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f525b59df454ccf53da6cb201e0aa8d09f52a2ad.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f57f84892e2a8496169b7406e63b0d4f5aa63aaf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f57f84892e2a8496169b7406e63b0d4f5aa63aaf.hip
deleted file mode 100644
index 82d07ba479902..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f57f84892e2a8496169b7406e63b0d4f5aa63aaf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f5803aadd93e33567aa6b23100ce4fbb6c040dd6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f5803aadd93e33567aa6b23100ce4fbb6c040dd6.hip
deleted file mode 100644
index 985fe72984365..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f5803aadd93e33567aa6b23100ce4fbb6c040dd6.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f5f1797f6b672a55476348571ce17645c8a62869.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f5f1797f6b672a55476348571ce17645c8a62869.hip
deleted file mode 100644
index a6173bb54fb58..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f5f1797f6b672a55476348571ce17645c8a62869.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6566441ac3074578cfe45758ba0583c0da0a5ab.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6566441ac3074578cfe45758ba0583c0da0a5ab.hip
deleted file mode 100644
index 67ddf754dafc0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6566441ac3074578cfe45758ba0583c0da0a5ab.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f672bf80a78885428b2c02e522426470653a7351.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f672bf80a78885428b2c02e522426470653a7351.hip
deleted file mode 100644
index 62da3abbfb5ee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f672bf80a78885428b2c02e522426470653a7351.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f682399cd6412fed6a1141296a7e4d42078f7b29.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f682399cd6412fed6a1141296a7e4d42078f7b29.hip
deleted file mode 100644
index 5a5c93702d7de..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f682399cd6412fed6a1141296a7e4d42078f7b29.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6856ca950bcf173571766c3f04de4163be0402e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6856ca950bcf173571766c3f04de4163be0402e.hip
deleted file mode 100644
index 45a2bd5fc9af6..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6856ca950bcf173571766c3f04de4163be0402e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f69548d6cced86c21c09c6475237a0cb926df0ed.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f69548d6cced86c21c09c6475237a0cb926df0ed.hip
deleted file mode 100644
index 1d2c930df99a7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f69548d6cced86c21c09c6475237a0cb926df0ed.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f69878f4ca8cfe6b8d8748766f66a1ef8eab20ad.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f69878f4ca8cfe6b8d8748766f66a1ef8eab20ad.hip
deleted file mode 100644
index 9c3ff824c9eee..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f69878f4ca8cfe6b8d8748766f66a1ef8eab20ad.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6f102a388ffb05c690a20a29cfe0b35a35eed61.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6f102a388ffb05c690a20a29cfe0b35a35eed61.hip
deleted file mode 100644
index 8d1d9830b17d1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f6f102a388ffb05c690a20a29cfe0b35a35eed61.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7035f4bfd8f2f427720a07e3c311bccc1dba683.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7035f4bfd8f2f427720a07e3c311bccc1dba683.hip
deleted file mode 100644
index 92b7645f8d661..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7035f4bfd8f2f427720a07e3c311bccc1dba683.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f71f96ce4dcc7f789a8ace73c230c203b05ff6dc.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f71f96ce4dcc7f789a8ace73c230c203b05ff6dc.hip
deleted file mode 100644
index b47a9b453bfca..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f71f96ce4dcc7f789a8ace73c230c203b05ff6dc.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f727911254904ce4341e4ff5f8bafc430b8cfbbf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f727911254904ce4341e4ff5f8bafc430b8cfbbf.hip
deleted file mode 100644
index f41a1ead45e69..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f727911254904ce4341e4ff5f8bafc430b8cfbbf.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::bf16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f731289837f915e2aec1bd01eef1b3c1b099864d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f731289837f915e2aec1bd01eef1b3c1b099864d.hip
deleted file mode 100644
index a530df9ff2fea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f731289837f915e2aec1bd01eef1b3c1b099864d.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, true, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        32,
-        false,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<32,
-                                                             ck_tile::bf16_t,
-                                                             false,
-                                                             true,
-                                                             true,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f79def2b4edf6d18f6ef1d6b141f9e0435441f6a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f79def2b4edf6d18f6ef1d6b141f9e0435441f6a.hip
deleted file mode 100644
index 50441617f1a14..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f79def2b4edf6d18f6ef1d6b141f9e0435441f6a.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, false,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7aa9c39b06e55bf4bc9f9a2a0fb075c9d4e69ce.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7aa9c39b06e55bf4bc9f9a2a0fb075c9d4e69ce.hip
deleted file mode 100644
index ee924640e3ec4..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7aa9c39b06e55bf4bc9f9a2a0fb075c9d4e69ce.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7cf08242b3fb1c643d4149bec985b667b9d28fa.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7cf08242b3fb1c643d4149bec985b667b9d28fa.hip
deleted file mode 100644
index 87dd684ce7072..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f7cf08242b3fb1c643d4149bec985b667b9d28fa.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f851da732f397624717160f89271514bc334b59b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f851da732f397624717160f89271514bc334b59b.hip
deleted file mode 100644
index 2d9cc06d4caf8..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f851da732f397624717160f89271514bc334b59b.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f861d8693f82d22e2c5b1abbcbae5f30f4433e5e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f861d8693f82d22e2c5b1abbcbae5f30f4433e5e.hip
deleted file mode 100644
index 105e0f9bde4f1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f861d8693f82d22e2c5b1abbcbae5f30f4433e5e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f87790f260630f312b84888dcbdf849ce130ae59.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f87790f260630f312b84888dcbdf849ce130ae59.hip
deleted file mode 100644
index 4abd9dc4272a9..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f87790f260630f312b84888dcbdf849ce130ae59.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f87991cb7787a29d3ce4711b4ce04c5fb6a14ca9.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f87991cb7787a29d3ce4711b4ce04c5fb6a14ca9.hip
deleted file mode 100644
index 83789c37fc346..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f87991cb7787a29d3ce4711b4ce04c5fb6a14ca9.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f90410c26d7649e21e2ae5e32e7af89d84d2ea70.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f90410c26d7649e21e2ae5e32e7af89d84d2ea70.hip
deleted file mode 100644
index 6959c135d2e42..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f90410c26d7649e21e2ae5e32e7af89d84d2ea70.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f92e9a82c879051d6fe3c42108f8a574187704af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f92e9a82c879051d6fe3c42108f8a574187704af.hip
deleted file mode 100644
index 3086388a0d6a2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f92e9a82c879051d6fe3c42108f8a574187704af.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f93bc23b8a4f1e0fc5c5756c4e1c835bf59dea09.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f93bc23b8a4f1e0fc5c5756c4e1c835bf59dea09.hip
deleted file mode 100644
index 9bac288ffe43a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f93bc23b8a4f1e0fc5c5756c4e1c835bf59dea09.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f93bf815b520a9d9e17b43bf9d7fb870751b6225.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f93bf815b520a9d9e17b43bf9d7fb870751b6225.hip
deleted file mode 100644
index 589eaecd9cffc..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f93bf815b520a9d9e17b43bf9d7fb870751b6225.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f974b12e83e214c30995a25631d37df1478927af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f974b12e83e214c30995a25631d37df1478927af.hip
deleted file mode 100644
index 0231413a7ac6b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f974b12e83e214c30995a25631d37df1478927af.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::fp16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f9824fb32933b27501ae8a7f43f460a2dda6a814.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f9824fb32933b27501ae8a7f43f460a2dda6a814.hip
deleted file mode 100644
index 0c37c9de80ccb..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f9824fb32933b27501ae8a7f43f460a2dda6a814.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f98a6b193fec3203eaa75819f6b51aa45a48f212.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f98a6b193fec3203eaa75819f6b51aa45a48f212.hip
deleted file mode 100644
index a0dc4114ef032..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f98a6b193fec3203eaa75819f6b51aa45a48f212.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f9c58761c927b222112cb5cb6c9acb5d3c915785.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f9c58761c927b222112cb5cb6c9acb5d3c915785.hip
deleted file mode 100644
index e14383b3e0b8a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_f9c58761c927b222112cb5cb6c9acb5d3c915785.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa16fa84278b489af253b52839786f94aeeac36f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa16fa84278b489af253b52839786f94aeeac36f.hip
deleted file mode 100644
index 314094b09d178..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa16fa84278b489af253b52839786f94aeeac36f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa62a97675719c2e8e9bb97361b92ff1c7b9d2ef.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa62a97675719c2e8e9bb97361b92ff1c7b9d2ef.hip
deleted file mode 100644
index c8195f0a1b63b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa62a97675719c2e8e9bb97361b92ff1c7b9d2ef.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa85f869a92f0482605e52019828244b12e12b44.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa85f869a92f0482605e52019828244b12e12b44.hip
deleted file mode 100644
index d0233e5328881..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fa85f869a92f0482605e52019828244b12e12b44.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fabdc143c29d5ca50ab1e96a814bda6d05b0d5d2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fabdc143c29d5ca50ab1e96a814bda6d05b0d5d2.hip
deleted file mode 100644
index b368189478363..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fabdc143c29d5ca50ab1e96a814bda6d05b0d5d2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fac5a0f98b94530befd634891e42c424bb86f0e1.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fac5a0f98b94530befd634891e42c424bb86f0e1.hip
deleted file mode 100644
index 67c384812c6d0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fac5a0f98b94530befd634891e42c424bb86f0e1.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fac99c3c82b77946f6844699d2333cd532a78a26.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fac99c3c82b77946f6844699d2333cd532a78a26.hip
deleted file mode 100644
index 38a8b5a1fffba..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fac99c3c82b77946f6844699d2333cd532a78a26.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 128, 32, 128>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<128, ck_tile::fp16_t, true,128, 128, 32, 128, 32, 128, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_faf56e45b2240515e97fc1bfd552eb03b6de5094.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_faf56e45b2240515e97fc1bfd552eb03b6de5094.hip
deleted file mode 100644
index 500e8708ed3f3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_faf56e45b2240515e97fc1bfd552eb03b6de5094.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_faf686067fa433cea5e95dd523846dc881eff635.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_faf686067fa433cea5e95dd523846dc881eff635.hip
deleted file mode 100644
index 9ce35b80def4f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_faf686067fa433cea5e95dd523846dc881eff635.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb2fbb135d59028afcf867c2cf08edc323565528.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb2fbb135d59028afcf867c2cf08edc323565528.hip
deleted file mode 100644
index e30a1d461cdbe..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb2fbb135d59028afcf867c2cf08edc323565528.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb4c15452f9155c5966990f09432e5eb7e28e785.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb4c15452f9155c5966990f09432e5eb7e28e785.hip
deleted file mode 100644
index 10d32f74fda04..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb4c15452f9155c5966990f09432e5eb7e28e785.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb4c5f8fecfbbe16e6648becb3b5ca89fa3d8a94.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb4c5f8fecfbbe16e6648becb3b5ca89fa3d8a94.hip
deleted file mode 100644
index 77fe95aa652e2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb4c5f8fecfbbe16e6648becb3b5ca89fa3d8a94.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb5bb49928ce5515d7b297d5eadd4ec70a22d60b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb5bb49928ce5515d7b297d5eadd4ec70a22d60b.hip
deleted file mode 100644
index 45a2509202342..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb5bb49928ce5515d7b297d5eadd4ec70a22d60b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb79e1f9231692d736dbada062ed6821f34927bf.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb79e1f9231692d736dbada062ed6821f34927bf.hip
deleted file mode 100644
index f1f09196bc698..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb79e1f9231692d736dbada062ed6821f34927bf.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb9477a613665cebcad781389ba7c5a36f51efe2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb9477a613665cebcad781389ba7c5a36f51efe2.hip
deleted file mode 100644
index 3bb8ba9f058ae..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fb9477a613665cebcad781389ba7c5a36f51efe2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fba36678d5047ded97ee7a7ba9feb9569afdb6ea.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fba36678d5047ded97ee7a7ba9feb9569afdb6ea.hip
deleted file mode 100644
index 53cc0538ac892..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fba36678d5047ded97ee7a7ba9feb9569afdb6ea.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fba47fa8d9b5375bc408af68b67345ab9dba2eb8.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fba47fa8d9b5375bc408af68b67345ab9dba2eb8.hip
deleted file mode 100644
index 55cabdff7bb28..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fba47fa8d9b5375bc408af68b67345ab9dba2eb8.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::bf16_t, false,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, false, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fbea85b766bf0c918ee0baf24dffc6a5563d5105.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fbea85b766bf0c918ee0baf24dffc6a5563d5105.hip
deleted file mode 100644
index 9921145904e62..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fbea85b766bf0c918ee0baf24dffc6a5563d5105.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fbeec221cd63adaedceec39db41ea942f99f5133.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fbeec221cd63adaedceec39db41ea942f99f5133.hip
deleted file mode 100644
index 09f49e4605439..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fbeec221cd63adaedceec39db41ea942f99f5133.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc030b61ae20c4b7d9b2d10930a17e01e9e93328.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc030b61ae20c4b7d9b2d10930a17e01e9e93328.hip
deleted file mode 100644
index d7cbb6a3f2f0a..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc030b61ae20c4b7d9b2d10930a17e01e9e93328.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc1790325b59bd44b0a5f6cf9723a25fd845cba7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc1790325b59bd44b0a5f6cf9723a25fd845cba7.hip
deleted file mode 100644
index 9a28018340c94..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc1790325b59bd44b0a5f6cf9723a25fd845cba7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc1eb85a00017efdc610e4259d2abe935b85304f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc1eb85a00017efdc610e4259d2abe935b85304f.hip
deleted file mode 100644
index 1a6edf526b175..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc1eb85a00017efdc610e4259d2abe935b85304f.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc5841a729099340d608e31023acbeaeade3e886.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc5841a729099340d608e31023acbeaeade3e886.hip
deleted file mode 100644
index c2246bfb6ac1b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc5841a729099340d608e31023acbeaeade3e886.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc5ebf0f2200f37ccc0849e0c3745f6e2f00111d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc5ebf0f2200f37ccc0849e0c3745f6e2f00111d.hip
deleted file mode 100644
index 856f565889e8c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc5ebf0f2200f37ccc0849e0c3745f6e2f00111d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc7b0916744b593435d8e1e7b6d874d760cd5e3b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc7b0916744b593435d8e1e7b6d874d760cd5e3b.hip
deleted file mode 100644
index 29d0b32e81bb1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc7b0916744b593435d8e1e7b6d874d760cd5e3b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc86c13e933cba40553ffba31d53aad27415ce4b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc86c13e933cba40553ffba31d53aad27415ce4b.hip
deleted file mode 100644
index e9dbee4392b2b..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fc86c13e933cba40553ffba31d53aad27415ce4b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcb0b08e29b2e1bf181fceceb9dc416e54f52b00.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcb0b08e29b2e1bf181fceceb9dc416e54f52b00.hip
deleted file mode 100644
index ab61511535132..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcb0b08e29b2e1bf181fceceb9dc416e54f52b00.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcb6ef39c3db49f26f736d6c9221dd825409ec4e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcb6ef39c3db49f26f736d6c9221dd825409ec4e.hip
deleted file mode 100644
index 43036d391d724..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcb6ef39c3db49f26f736d6c9221dd825409ec4e.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    true,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, true,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, true, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcbe827108d252b2f5847fa8e132c9c3e56a90a0.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcbe827108d252b2f5847fa8e132c9c3e56a90a0.hip
deleted file mode 100644
index afbe0fc50eb6f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fcbe827108d252b2f5847fa8e132c9c3e56a90a0.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fccabea88b8e290688c1b360875d228e6fdf1624.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fccabea88b8e290688c1b360875d228e6fdf1624.hip
deleted file mode 100644
index efeeb811fb30c..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fccabea88b8e290688c1b360875d228e6fdf1624.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd10a3b937e9659716925e39a01d794914b08e26.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd10a3b937e9659716925e39a01d794914b08e26.hip
deleted file mode 100644
index dce14072009cd..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd10a3b937e9659716925e39a01d794914b08e26.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 16, 32, 32, 32>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<2, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    true,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::fp16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::fp16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_HBS<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<32, ck_tile::fp16_t, true,128, 64, 16, 32, 32, 32, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, true, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd19d7614f2ed5da21a52ed172ef62cc07c9c01a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd19d7614f2ed5da21a52ed172ef62cc07c9c01a.hip
deleted file mode 100644
index 56fcf8377dbea..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd19d7614f2ed5da21a52ed172ef62cc07c9c01a.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        true,
-        false,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::bf16_t,
-                                                             true,
-                                                             true,
-                                                             false,
-                                                             false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd26e43ca652e6f58ff48c356165aa4349833b55.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd26e43ca652e6f58ff48c356165aa4349833b55.hip
deleted file mode 100644
index c81a859cd7c02..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd26e43ca652e6f58ff48c356165aa4349833b55.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd345632e0cae0d549ba79626a08b1885711deb6.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd345632e0cae0d549ba79626a08b1885711deb6.hip
deleted file mode 100644
index 3dca81219aff1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd345632e0cae0d549ba79626a08b1885711deb6.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 128, 32, 256, 32, 256>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<true>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVS<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<256, ck_tile::bf16_t, false,128, 128, 32, 256, 32, 256, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::NO_BIAS, false, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd3558b4c7a667dbc365c4c2ceda646975408f51.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd3558b4c7a667dbc365c4c2ceda646975408f51.hip
deleted file mode 100644
index 8d236cf911d0d..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd3558b4c7a667dbc365c4c2ceda646975408f51.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd614df484b263deae3b3c20adb0ce7b62eaa651.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd614df484b263deae3b3c20adb0ce7b62eaa651.hip
deleted file mode 100644
index 9de1995c234f3..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd614df484b263deae3b3c20adb0ce7b62eaa651.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd9cd1305633b62b68fb8474ce021f639f8492e7.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd9cd1305633b62b68fb8474ce021f639f8492e7.hip
deleted file mode 100644
index abbe42c081a35..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fd9cd1305633b62b68fb8474ce021f639f8492e7.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fde12cd366d6850ce26afce98e5076b695b4875b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fde12cd366d6850ce26afce98e5076b695b4875b.hip
deleted file mode 100644
index 1203456f8d9c2..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fde12cd366d6850ce26afce98e5076b695b4875b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe245e9ea974adce2b9807d33b9ba12d916eaffb.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe245e9ea974adce2b9807d33b9ba12d916eaffb.hip
deleted file mode 100644
index e4e3c95292d43..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe245e9ea974adce2b9807d33b9ba12d916eaffb.hip
+++ /dev/null
@@ -1,80 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_fwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::sequence<128, 64, 32, 64, 32, 64>;
-using fmha_warp_tile_0 = ck_tile::sequence<32, 32, 16>;
-
-using fmha_shape_0 = ck_tile::TileFmhaShape<fmha_block_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      ck_tile::sequence<4, 1, 1>,
-                                      fmha_warp_tile_0,
-                                      true>;
-
-using fmha_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                    true,
-                                                    true,
-                                                    true,
-                                                    ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                    false,
-                                                    true,
-                                                    false,
-                                                    false,
-                                                    -1>;
-using fmha_mask_0 = ck_tile::SimplifiedGenericAttentionMask<false>;
-
-using fmha_pipeline_problem_0 = ck_tile::BlockFmhaPipelineProblem<
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::SMPLComputeDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::PDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::OaccDataType,
-    typename FmhaFwdTypeConfig<fmha_dtype_0>::ODataType,
-    fmha_shape_0,
-    false,
-    fmha_mask_0,
-    fmha_trait_0>;
-
-using fmha_pipeline_0 = ck_tile::BlockFmhaPipelineQRKSVSAsync<
-    fmha_pipeline_problem_0>;
-
-using fmha_epilogue_0 =
-    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<ck_tile::bf16_t>::OaccDataType,
-                                           typename FmhaFwdTypeConfig<ck_tile::bf16_t>::ODataType,
-                                           true, true>>;
-
-using fmha_kernel_0 =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner_SHB<fmha_shape_0>,
-                  fmha_pipeline_0,
-                  fmha_epilogue_0>;
-
-using trait_0 = fmha_fwd_traits_<64, ck_tile::bf16_t, false,128, 64, 32, 64, 32, 64, true,
-                        ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC, fmha_mask_0, ck_tile::BlockAttentionBiasEnum::ALIBI, true, false, false, true, true, true, true>;
-
-#include <iostream>
-
-template<>
-float fmha_fwd_<trait_0>(const ck_tile::stream_config& s, fmha_fwd_args a)
-{
-    using k_ = fmha_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks             = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe72cdd69944d2d765478d4aed13066a02b76f6d.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe72cdd69944d2d765478d4aed13066a02b76f6d.hip
deleted file mode 100644
index 0985b165254e5..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe72cdd69944d2d765478d4aed13066a02b76f6d.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 64, 32, 64, 32, 32, 64, 64>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<64,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe8b8c3525fe86a20a2d6c69585f3e36c16caabd.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe8b8c3525fe86a20a2d6c69585f3e36c16caabd.hip
deleted file mode 100644
index dcafc647b4813..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe8b8c3525fe86a20a2d6c69585f3e36c16caabd.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    true,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         true,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe97b7adcd67ed9bda8831d1f3f1ca7590c6d251.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe97b7adcd67ed9bda8831d1f3f1ca7590c6d251.hip
deleted file mode 100644
index a904086533b86..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe97b7adcd67ed9bda8831d1f3f1ca7590c6d251.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe9d98dbec5096a89b116f85675af772f023014a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe9d98dbec5096a89b116f85675af772f023014a.hip
deleted file mode 100644
index 7cf999339a077..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fe9d98dbec5096a89b116f85675af772f023014a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_feb5e77111fe1e20bafdb83a925b5faeeb6214af.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_feb5e77111fe1e20bafdb83a925b5faeeb6214af.hip
deleted file mode 100644
index 88d22e1537eff..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_feb5e77111fe1e20bafdb83a925b5faeeb6214af.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fecd7501265b4c4dcf015485e63e2324304f70d3.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fecd7501265b4c4dcf015485e63e2324304f70d3.hip
deleted file mode 100644
index 7c1ee5f4a10da..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fecd7501265b4c4dcf015485e63e2324304f70d3.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fecffa403b3631b1957e1a9a06f18fdb3b4eee5f.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fecffa403b3631b1957e1a9a06f18fdb3b4eee5f.hip
deleted file mode 100644
index 3c233e8412a8f..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fecffa403b3631b1957e1a9a06f18fdb3b4eee5f.hip
+++ /dev/null
@@ -1,65 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_bwd_dot_do_o_trait_0 =
-    ck_tile::TileFmhaBwdOGradDotOTraits<true, true, 2>;
-
-using fmha_bwd_dot_do_o_pipeline_problem_0 = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    /* BlockSize = */ 64,
-    32,
-    false,
-    fmha_bwd_dot_do_o_trait_0>;
-
-using fmha_bwd_dot_do_o_0 =
-    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_0>;
-
-using fmha_bwd_dot_do_o_kernel_0 =
-    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_0>;
-
-using dot_do_o_trait_0 =
-    fmha_bwd_dot_do_o_traits_<32, ck_tile::bf16_t, false, true, true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dot_do_o_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dot_do_o_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_0>()
-{
-    using k_ = fmha_bwd_dot_do_o_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ff453e3bdc9752cb7b81f7cc3056325a8b9a8ad4.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ff453e3bdc9752cb7b81f7cc3056325a8b9a8ad4.hip
deleted file mode 100644
index 3ad7535c5405e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ff453e3bdc9752cb7b81f7cc3056325a8b9a8ad4.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ff6862dbdbb20bc63a650e1f93e9ac169bb702b2.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ff6862dbdbb20bc63a650e1f93e9ac169bb702b2.hip
deleted file mode 100644
index 658f8e0f01b37..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ff6862dbdbb20bc63a650e1f93e9ac169bb702b2.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffb5b7349a671b182d73c8016590f26fe06a4cba.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffb5b7349a671b182d73c8016590f26fe06a4cba.hip
deleted file mode 100644
index c71bd6afbbfc7..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffb5b7349a671b182d73c8016590f26fe06a4cba.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<true,
-                                                       false,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      false,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffb8adef0cef91a86f36872407fea35df90e8f2b.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffb8adef0cef91a86f36872407fea35df90e8f2b.hip
deleted file mode 100644
index 3e48b55ce5c88..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffb8adef0cef91a86f36872407fea35df90e8f2b.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::bf16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::bf16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::bf16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::bf16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffc6056d9fe125a4dbe08c1d86354e51f7daadd5.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffc6056d9fe125a4dbe08c1d86354e51f7daadd5.hip
deleted file mode 100644
index 3d71b992a99d1..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffc6056d9fe125a4dbe08c1d86354e51f7daadd5.hip
+++ /dev/null
@@ -1,73 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_bwd_convert_dq_trait_0 =
-    ck_tile::TileFmhaBwdConvertQGradTraits<true, false, 2>;
-
-using fmha_bwd_convert_dq_pipeline_problem_0 =
-    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-        typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-        /* BlockSize = */ 256,
-        64,
-        128,
-        64,
-        false,
-        true,
-        fmha_bwd_convert_dq_trait_0>;
-
-using fmha_bwd_convert_dq_0 =
-    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_0>;
-
-using fmha_bwd_convert_dq_kernel_0 =
-    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_0>;
-
-using convert_dq_trait_0 = fmha_bwd_convert_dq_traits_<64,
-                                                             ck_tile::fp16_t,
-                                                             false,
-                                                             true,
-                                                             false,
-                                                             true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_convert_dq_<convert_dq_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_0>(const ck_tile::stream_config& s,
-                                                            fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_convert_dq_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_0>()
-{
-    using k_ = fmha_bwd_convert_dq_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffd868d49abdb769ab82c21508d655daf54b8a99.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffd868d49abdb769ab82c21508d655daf54b8a99.hip
deleted file mode 100644
index c95405c962a78..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_ffd868d49abdb769ab82c21508d655daf54b8a99.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 64, 256, 16, 256, 16, 32, 256, 256>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       true,
-                                                       true,
-                                                       ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<false, true,  false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      true>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<256,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::NO_BIAS,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         true,
-                                                         true,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fff7aa57cca501f221077124359a589b3a6f9d0a.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fff7aa57cca501f221077124359a589b3a6f9d0a.hip
deleted file mode 100644
index 3822d9fcbd710..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fff7aa57cca501f221077124359a589b3a6f9d0a.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<16, 128, 128, 16, 128, 16, 32, 128, 128>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<true>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    true,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      false,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<128,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         false,
-                                                         true>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fffbfcac254e33926131a71905e93f9cc0aef89e.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fffbfcac254e33926131a71905e93f9cc0aef89e.hip
deleted file mode 100644
index 56f1ea131d359..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_ck_autogen_fffbfcac254e33926131a71905e93f9cc0aef89e.hip
+++ /dev/null
@@ -1,138 +0,0 @@
-// ==========================================
-// THIS CODE IS AUTOGENERATED. DO NOT MODIFY.
-// @generated
-// ==========================================
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-// auto generated by generate.py
-#include <fmha_bwd.hpp>
-
-using fmha_dtype_0 = ck_tile::fp16_t;
-
-using fmha_block_tile_0 = ck_tile::
-    sequence<32, 128, 32, 32, 32, 32, 64, 32, 32>;
-using fmha_block_warps0_0 = ck_tile::sequence<1, 4, 1>;
-using fmha_block_warps1_0 = ck_tile::sequence<4, 1, 1>;
-using fmha_block_warps2_0 = ck_tile::sequence<2, 2, 1>;
-using fmha_warp_tile0_0   = ck_tile::sequence<16, 16, 32>;
-using fmha_warp_tile1_0   = ck_tile::sequence<16, 16, 16>;
-
-// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
-//       G0&G2 -> GSdP
-//       G1&G3 -> GdKV
-//       G4    -> GdQ
-using fmha_bwd_shape_0 = ck_tile::TileFmhaBwdShape<fmha_block_tile_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps0_0,
-                                                         fmha_warp_tile0_0,
-                                                         fmha_block_warps1_0,
-                                                         fmha_warp_tile1_0,
-                                                         fmha_block_warps2_0,
-                                                         fmha_warp_tile0_0>;
-
-using fmha_bwd_trait_0 = ck_tile::TileFmhaTraits<false,
-                                                       true,
-                                                       false,
-                                                       false,
-                                                       ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       false,
-                                                       1>;
-using fmha_mask_0      = ck_tile::SimplifiedGenericAttentionMask<false>;
-using fmha_dropout_0   = ck_tile::BlockDropoutBwd<true,  false, false>;
-
-using fmha_bwd_pipeline_problem_0 = ck_tile::BlockFmhaBwdPipelineProblem<
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::GemmDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::LSEDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::AccDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::DDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::RandValOutputDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::ODataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::OGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::QGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::KGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::VGradDataType,
-    typename FmhaBwdTypeConfig<fmha_dtype_0>::BiasGradDataType,
-    fmha_bwd_shape_0,
-    false,
-    false,
-    fmha_mask_0,
-    fmha_dropout_0,
-    fmha_bwd_trait_0>;
-
-using fmha_bwd_pipeline_0 = ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP<fmha_bwd_pipeline_problem_0>;
-
-using fmha_bwd_dk_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::KGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dv_epilogue_0 = ck_tile::Default2DEpilogue<
-    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<ck_tile::fp16_t>::AccDataType,
-                                      typename FmhaBwdTypeConfig<ck_tile::fp16_t>::VGradDataType,
-                                      true,
-                                      false>>;
-
-using fmha_bwd_dq_dk_dv_kernel_0 =
-    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_0,
-                                 fmha_bwd_dk_epilogue_0,
-                                 fmha_bwd_dv_epilogue_0>;
-
-using dq_dk_dv_trait_0 = fmha_bwd_dq_dk_dv_traits_<32,
-                                                         ck_tile::fp16_t,
-                                                         false,
-                                                         ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP,
-                                                         fmha_mask_0,
-                                                         fmha_dropout_0,
-                                                         ck_tile::BlockAttentionBiasEnum::ALIBI,
-                                                         false,
-                                                         false,
-                                                         true,
-                                                         false,
-                                                         false,
-                                                         false>;
-
-#include <iostream>
-
-template <>
-float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s, fmha_bwd_args a)
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    if(s.log_level_ > 0)
-        std::cout << ", " << k_::GetName() << std::flush;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel(
-        s, ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs));
-}
-
-template <>
-void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_0>(const ck_tile::stream_config& s,
-                                                        fmha_bwd_args a)
-{
-    using k_                               = fmha_bwd_dq_dk_dv_kernel_0;
-    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
-    constexpr dim3 blocks                  = k_::BlockSize();
-    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    ck_tile::make_kernel_pt<blocks.x, kBlockPerCu>(k_{}, grids, blocks, 0, kargs)(
-        ck_tile::stream_config{s.stream_id_});
-}
-
-template <>
-std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_0>()
-{
-    using k_ = fmha_bwd_dq_dk_dv_kernel_0;
-    return k_::GetName();
-}
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_fwd.hpp b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_fwd.hpp
index d4ccb668c6028..2de70cd49bbb7 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_fwd.hpp
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/fmha_fwd.hpp
@@ -17,11 +17,35 @@
 #include <utility>
 #include <variant>
 
+struct FmhaFwdFp16
+{
+};
+
+struct FmhaFwdBf16
+{
+};
+
+struct FmhaFwdFp8
+{
+};
+
+struct FmhaFwdBf8
+{
+};
+
+struct FmhaFwdFp8Fp16
+{
+};
+
+struct FmhaFwdFp8Bf16
+{
+};
+
 template <typename DataType>
 struct FmhaFwdTypeConfig;
 
 template <>
-struct FmhaFwdTypeConfig<ck_tile::half_t>
+struct FmhaFwdTypeConfig<FmhaFwdFp16>
 {
     using QDataType             = ck_tile::half_t;
     using KDataType             = ck_tile::half_t;
@@ -37,7 +61,7 @@ struct FmhaFwdTypeConfig<ck_tile::half_t>
 };
 
 template <>
-struct FmhaFwdTypeConfig<ck_tile::bf16_t>
+struct FmhaFwdTypeConfig<FmhaFwdBf16>
 {
     using QDataType             = ck_tile::bf16_t;
     using KDataType             = ck_tile::bf16_t;
@@ -53,7 +77,7 @@ struct FmhaFwdTypeConfig<ck_tile::bf16_t>
 };
 
 template <>
-struct FmhaFwdTypeConfig<ck_tile::fp8_t>
+struct FmhaFwdTypeConfig<FmhaFwdFp8>
 {
     using QDataType             = ck_tile::fp8_t;
     using KDataType             = ck_tile::fp8_t;
@@ -69,7 +93,7 @@ struct FmhaFwdTypeConfig<ck_tile::fp8_t>
 };
 
 template <>
-struct FmhaFwdTypeConfig<ck_tile::bf8_t>
+struct FmhaFwdTypeConfig<FmhaFwdBf8>
 {
     using QDataType             = ck_tile::bf8_t;
     using KDataType             = ck_tile::bf8_t;
@@ -166,6 +190,8 @@ struct fmha_fwd_splitkv_args
     void* block_table_ptr;
     ck_tile::index_t batch_stride_block_table; // only used if 'block_table_ptr' is not nullptr
     ck_tile::index_t page_block_size;          // only used if 'block_table_ptr' is not nullptr
+    bool is_gappy; // differentiate seqstart_k_ptr usage. only used if 'block_table_ptr' is not
+                   // nullptr.
 
     const void* cache_batch_idx;
 
@@ -174,9 +200,21 @@ struct fmha_fwd_splitkv_args
     //             seqlen_k = kargs.seqlen_k
     // group mode: seqlen_q = kargs.seqstart_q_ptr[b + 1] - kargs.seqstart_q_ptr[b]
     //             seqlen_k = kargs.seqstart_k_ptr[b + 1] - kargs.seqstart_k_ptr[b]
-    // kvcache mode (use same kernel as batch mode):
+    //                      or kargs.seqlen_k_ptr[b]
+    //
+    // batch mode (kvcache):
     //             seqlen_q = kargs.seqlen_q
+    //             seqlen_k = kargs.seqlen_k_ptr[b]
+    // group mode (kvcache):
+    //             seqlen_q = kargs.seqstart_q_ptr[b + 1] - kargs.seqstart_q_ptr[b]
+    //
+    //     when is_gappy=true:
+    //             seqlen_k = kargs.seqlen_k_ptr[b]
+    //             seqstart_k_ptr[b] now store local offset of each batch
+    //
+    //     when is_gappy=false:
     //             seqlen_k = kargs.seqstart_k_ptr[b + 1] - kargs.seqstart_k_ptr[b]
+    //                      or kargs.seqlen_k_ptr[b]
     const void* seqstart_q_ptr;
     const void* seqstart_k_ptr;
     const void* seqlen_k_ptr;
@@ -252,7 +290,7 @@ struct fmha_fwd_appendkv_args
     ck_tile::index_t batch_stride_block_table; // only used if 'block_table_ptr' is not nullptr
     ck_tile::index_t page_block_size;          // only used if 'block_table_ptr' is not nullptr
 
-    const void* cache_batch_idx;
+    const void* cache_batch_idx; // only used if block_table_ptr is nullptr -> batch mode (kvcache)
 
     ck_tile::index_t stride_q;
     ck_tile::index_t stride_k;
@@ -280,91 +318,101 @@ auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
         if constexpr(FmhaKernel::kIsGroupMode)
         {
             return FmhaKernel::MakeKargsImpl(args.q_ptr,
-                                         args.k_ptr,
-                                         args.v_ptr,
-                                         args.bias_ptr,
-                                         args.rand_val_ptr,
-                                         args.lse_ptr,
-                                         args.o_ptr,
-                                         args.seqstart_q_ptr,
-                                         args.seqstart_k_ptr,
-                                         args.seqlen_k_ptr,
-                                         args.hdim_q,
-                                         args.hdim_v,
-                                         args.nhead_q,
-                                         args.nhead_q / args.nhead_k,
-                                         args.scale_s,
-                                         args.scale_p,
-                                         args.scale_o,
-                                         args.stride_q,
-                                         args.stride_k,
-                                         args.stride_v,
-                                         args.stride_bias,
-                                         args.stride_randval,
-                                         args.stride_o,
-                                         args.nhead_stride_q,
-                                         args.nhead_stride_k,
-                                         args.nhead_stride_v,
-                                         args.nhead_stride_bias,
-                                         args.nhead_stride_randval,
-                                         args.nhead_stride_lse,
-                                         args.nhead_stride_o,
-                                         args.window_size_left,
-                                         args.window_size_right,
-                                         args.mask_type,
-                                         args.p_drop,
-                                         args.s_randval,
-                                         args.drop_seed_offset);
+                                             args.k_ptr,
+                                             args.v_ptr,
+                                             args.bias_ptr,
+                                             args.rand_val_ptr,
+                                             args.lse_ptr,
+                                             args.o_ptr,
+                                             args.seqstart_q_ptr,
+                                             args.seqstart_k_ptr,
+                                             args.seqlen_k_ptr,
+                                             args.hdim_q,
+                                             args.hdim_v,
+                                             args.nhead_q,
+                                             args.nhead_q / args.nhead_k,
+                                             args.scale_s,
+                                             args.scale_p,
+                                             args.scale_o,
+                                             args.stride_q,
+                                             args.stride_k,
+                                             args.stride_v,
+                                             args.stride_bias,
+                                             args.stride_randval,
+                                             args.stride_o,
+                                             args.nhead_stride_q,
+                                             args.nhead_stride_k,
+                                             args.nhead_stride_v,
+                                             args.nhead_stride_bias,
+                                             args.nhead_stride_randval,
+                                             args.nhead_stride_lse,
+                                             args.nhead_stride_o,
+                                             args.window_size_left,
+                                             args.window_size_right,
+                                             args.mask_type,
+                                             args.p_drop,
+                                             args.s_randval,
+                                             args.drop_seed_offset);
         }
         else
         { // create batch mode kernel arguments
             return FmhaKernel::MakeKargsImpl(args.q_ptr,
-                                         args.k_ptr,
-                                         args.v_ptr,
-                                         args.bias_ptr,
-                                         args.rand_val_ptr,
-                                         args.lse_ptr,
-                                         args.o_ptr,
-                                         args.seqlen_q,
-                                         args.seqlen_k,
-                                         args.hdim_q,
-                                         args.hdim_v,
-                                         args.nhead_q,
-                                         args.nhead_q / args.nhead_k,
-                                         args.scale_s,
-                                         args.scale_p,
-                                         args.scale_o,
-                                         args.stride_q,
-                                         args.stride_k,
-                                         args.stride_v,
-                                         args.stride_bias,
-                                         args.stride_randval,
-                                         args.stride_o,
-                                         args.nhead_stride_q,
-                                         args.nhead_stride_k,
-                                         args.nhead_stride_v,
-                                         args.nhead_stride_bias,
-                                         args.nhead_stride_randval,
-                                         args.nhead_stride_lse,
-                                         args.nhead_stride_o,
-                                         args.batch_stride_q,
-                                         args.batch_stride_k,
-                                         args.batch_stride_v,
-                                         args.batch_stride_bias,
-                                         args.batch_stride_randval,
-                                         args.batch_stride_lse,
-                                         args.batch_stride_o,
-                                         args.window_size_left,
-                                         args.window_size_right,
-                                         args.mask_type,
-                                         args.p_drop,
-                                         args.s_randval,
-                                         args.drop_seed_offset);
+                                             args.k_ptr,
+                                             args.v_ptr,
+                                             args.bias_ptr,
+                                             args.rand_val_ptr,
+                                             args.lse_ptr,
+                                             args.o_ptr,
+                                             args.seqlen_q,
+                                             args.seqlen_k,
+                                             args.hdim_q,
+                                             args.hdim_v,
+                                             args.nhead_q,
+                                             args.nhead_q / args.nhead_k,
+                                             args.scale_s,
+                                             args.scale_p,
+                                             args.scale_o,
+                                             args.stride_q,
+                                             args.stride_k,
+                                             args.stride_v,
+                                             args.stride_bias,
+                                             args.stride_randval,
+                                             args.stride_o,
+                                             args.nhead_stride_q,
+                                             args.nhead_stride_k,
+                                             args.nhead_stride_v,
+                                             args.nhead_stride_bias,
+                                             args.nhead_stride_randval,
+                                             args.nhead_stride_lse,
+                                             args.nhead_stride_o,
+                                             args.batch_stride_q,
+                                             args.batch_stride_k,
+                                             args.batch_stride_v,
+                                             args.batch_stride_bias,
+                                             args.batch_stride_randval,
+                                             args.batch_stride_lse,
+                                             args.batch_stride_o,
+                                             args.window_size_left,
+                                             args.window_size_right,
+                                             args.mask_type,
+                                             args.p_drop,
+                                             args.s_randval,
+                                             args.drop_seed_offset);
         }
     }();
 
-    dim3 grids = FmhaKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v);
-    return ck_tile::make_tuple(kargs, grids);
+    if constexpr(FmhaKernel::kIsGroupMode)
+    {
+        dim3 grids = FmhaKernel::GridSize(
+            args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, args.seqlen_k_ptr != nullptr);
+        return ck_tile::make_tuple(kargs, grids);
+    }
+    else
+    {
+        dim3 grids =
+            FmhaKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, false);
+        return ck_tile::make_tuple(kargs, grids);
+    }
 }
 
 template <typename Kernel>
@@ -375,7 +423,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
         // create group mode kernel arguments
         if constexpr(Kernel::kIsGroupMode)
         {
-            return Kernel::MakeKargsImpl(args.q_ptr,
+            return Kernel::MakeKargs(args.q_ptr,
                                      args.k_ptr,
                                      args.v_ptr,
                                      args.bias_ptr,
@@ -390,6 +438,10 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
                                      args.nhead_q,
                                      args.nhead_q / args.nhead_k,
                                      args.num_splits,
+                                     args.block_table_ptr,
+                                     args.batch_stride_block_table,
+                                     args.page_block_size,
+                                     args.is_gappy,
                                      args.scale_s,
                                      args.scale_p,
                                      args.stride_q,
@@ -413,7 +465,7 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
         }
         else
         { // create batch mode kernel arguments
-            return Kernel::MakeKargsImpl(args.q_ptr,
+            return Kernel::MakeKargs(args.q_ptr,
                                      args.k_ptr,
                                      args.v_ptr,
                                      args.bias_ptr,
@@ -459,8 +511,8 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
         }
     }();
 
-    dim3 grids =
-        Kernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, args.num_splits);
+    dim3 grids = Kernel::GridSize(
+        args.batch, args.nhead_q, args.nhead_k, args.max_seqlen_q, args.hdim_v, args.num_splits);
 
     return ck_tile::make_tuple(kargs, grids);
 }
@@ -473,7 +525,7 @@ auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_splitkv_args args)
         // create group mode kernel argumentszs
         if constexpr(Kernel::kIsGroupMode)
         {
-            return Kernel::MakeKargsImpl(args.lse_acc_ptr,
+            return Kernel::MakeKargs(args.lse_acc_ptr,
                                      args.o_acc_ptr,
                                      args.lse_ptr,
                                      args.o_ptr,
@@ -493,7 +545,7 @@ auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_splitkv_args args)
         }
         else
         { // create batch mode kernel arguments
-            return Kernel::MakeKargsImpl(args.lse_acc_ptr,
+            return Kernel::MakeKargs(args.lse_acc_ptr,
                                      args.o_acc_ptr,
                                      args.lse_ptr,
                                      args.o_ptr,
@@ -526,7 +578,7 @@ template <typename Kernel>
 auto fmha_fwd_appendkv_create_kargs_and_grids(fmha_fwd_appendkv_args args)
 {
     assert(args.nhead_q % args.nhead_k == 0);
-    auto kargs = Kernel::MakeKargsImpl(args.q_ptr,
+    auto kargs = Kernel::MakeKargs(args.q_ptr,
                                    args.k_ptr,
                                    args.knew_ptr,
                                    args.v_ptr,
@@ -668,7 +720,6 @@ std::string fmha_fwd_splitkv_get_name_();
 template <ck_tile::index_t HDim_,
           typename DataType_,
           bool kIsGroupMode_,
-          ck_tile::index_t kM0_,
           ck_tile::index_t kN1_,
           bool kStoreLse_,
           bool kDoFp8StaticQuant_,
@@ -679,7 +730,6 @@ struct fmha_fwd_splitkv_combine_traits_
     static constexpr ck_tile::index_t HDim  = HDim_;
     using DataType                          = ck_tile::remove_cvref_t<DataType_>;
     static constexpr bool kIsGroupMode      = kIsGroupMode_;
-    static constexpr ck_tile::index_t kM0   = kM0_;
     static constexpr ck_tile::index_t kN1   = kN1_;
     static constexpr bool kStoreLse         = kStoreLse_;
     static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_bwd_ck.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_bwd_ck.hip
new file mode 100644
index 0000000000000..b3ab46704f730
--- /dev/null
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_bwd_ck.hip
@@ -0,0 +1,124 @@
+#include <ATen/native/transformers/hip/flash_attn/flash_common_hip.hpp>
+#include <ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h>
+
+#include <iostream>
+
+namespace pytorch_flash {
+std::tuple<
+    at::Tensor, // dQ
+    at::Tensor, // dK
+    at::Tensor, // dV
+    at::Tensor> // dBias
+mem_eff_backward_ck(
+    const at::Tensor &dout,
+    const at::Tensor &q,
+    const at::Tensor &k,
+    const at::Tensor &v,
+    const at::Tensor &out,
+    const at::Tensor &softmax_lse,
+    const at::Tensor &dq_,
+    const at::Tensor &dk_,
+    const at::Tensor &dv_,
+    std::optional<at::Tensor> &attn_bias,
+    bool bias_requires_grad,
+    std::optional<at::Tensor> &grad_bias,
+    std::optional<at::Tensor> &cu_seqlens_q,
+    std::optional<at::Tensor> &cu_seqlens_k,
+    int max_seqlen_q,
+    int max_seqlen_k,
+    float p_dropout,
+    float scale,
+    bool is_causal,
+    bool deterministic,
+    bool zero_tensors,
+    at::Tensor philox_seed,
+    at::Tensor philox_offset)
+{
+// TODO implement wrapper
+  std::cout << "HIT MY MEM_EFF BWD ENTRY POINT" << std::endl;
+
+  const int non_null_window_left  = -1;
+  const int non_null_window_right = -1;
+
+  // Wrap gradients in std::optional
+  std::optional<at::Tensor> opt_dQ, opt_dK, opt_dV;
+  opt_dQ = dq_;
+  opt_dK = dk_;
+  opt_dV = dv_;
+
+  if(!cu_seqlens_q.has_value()) {
+    //  both of these return dq, dk, dv, softmax_d
+    // need to also return attn_bias
+    // call mha_bwd_ck
+    auto
+      [dQ,
+       dK,
+       dV,
+       softmax_d,
+       dBias] =
+        mha_bwd_ck(
+          dout,
+          q,
+          k,
+          v,
+          out,
+          softmax_lse,
+          opt_dQ,
+          opt_dK,
+          opt_dV,
+          attn_bias,
+          bias_requires_grad,
+          grad_bias,
+          p_dropout,
+          scale,
+          is_causal,
+          non_null_window_left,
+          non_null_window_right,
+          deterministic,
+          philox_seed,
+          philox_offset);
+    return std::make_tuple(std::move(dQ), std::move(dK), std::move(dV), std::move(dBias));
+
+  } else {
+    // call mha_varlen_bwd_ck
+    auto
+      [dQ,
+       dK,
+       dV,
+       softmax_d,
+       dBias] =
+        mha_varlen_bwd_ck(
+          dout,
+          q,
+          k,
+          v,
+          out,
+          softmax_lse,
+          opt_dQ,
+          opt_dK,
+          opt_dV,
+          cu_seqlens_q.value(),
+          cu_seqlens_k.value(),
+          attn_bias,
+          bias_requires_grad,
+          grad_bias,
+          max_seqlen_q,
+          max_seqlen_k,
+          p_dropout,
+          scale,
+          zero_tensors,
+          is_causal,
+          non_null_window_left,
+          non_null_window_right,
+          deterministic,
+          philox_seed,
+          philox_offset);
+    return std::make_tuple(std::move(dQ), std::move(dK), std::move(dV), std::move(dBias));
+
+  }
+
+  return std::make_tuple(at::Tensor{}, at::Tensor{}, at::Tensor{}, at::Tensor{});
+}
+
+
+} // namespace pytorch_flash
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h b/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h
new file mode 100644
index 0000000000000..5a284c7ac6b72
--- /dev/null
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h
@@ -0,0 +1,67 @@
+#pragma once
+#include <cstddef>
+
+#include <ATen/core/Tensor.h>
+
+
+namespace pytorch_flash {
+
+std::tuple<
+    at::Tensor, // output
+    at::Tensor, // q
+    at::Tensor, // k
+    at::Tensor, // v
+    at::Tensor, // lse
+    at::Tensor, // seed
+    at::Tensor, // offset
+    at::Tensor> // dropout randval
+mem_eff_forward_ck(
+    const at::Tensor& q,
+    const at::Tensor& k,
+    const at::Tensor& v,
+    float p_dropout,
+    bool return_dropout_randval,
+    std::optional<bool> is_causal,
+    std::optional<float> scale,
+    const std::optional<at::Tensor>& attn_bias_,
+    std::optional<at::Tensor>& out_,
+    const std::optional<at::Tensor>& cu_seqlens_q,
+    const std::optional<at::Tensor>& cu_seqlens_k,
+    const std::optional<at::Tensor>& seqstart_q,
+    const std::optional<at::Tensor>& seqstart_k,
+    std::optional<at::Generator> gen_,
+    std::optional<at::Tensor>& seqused_k_
+);
+
+// TODO get return tensors correct
+std::tuple<
+    at::Tensor, // dQ
+    at::Tensor, // dK
+    at::Tensor, // dV
+    at::Tensor> // dBias
+mem_eff_backward_ck(
+    const at::Tensor &dout,
+    const at::Tensor &q,
+    const at::Tensor &k,
+    const at::Tensor &v,
+    const at::Tensor &out,
+    const at::Tensor &softmax_lse,
+    const at::Tensor &dq_,
+    const at::Tensor &dk_,
+    const at::Tensor &dv_,
+    std::optional<at::Tensor> &attn_bias,
+    bool bias_requires_grad,
+    std::optional<at::Tensor> &grad_bias,
+    std::optional<at::Tensor> &cu_seqlens_q,
+    std::optional<at::Tensor> &cu_seqlens_k,
+    int max_seqlen_q,
+    int max_seqlen_k,
+    float p_dropout,
+    float scale,
+    bool is_causal,
+    bool deterministic,
+    bool zero_tensors,
+    const at::Tensor philox_seed,
+    const at::Tensor philox_offset);
+
+} // namespace pytorch_flash
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_fwd_ck.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_fwd_ck.hip
new file mode 100644
index 0000000000000..ae2652972cce7
--- /dev/null
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/me_fwd_ck.hip
@@ -0,0 +1,88 @@
+#include <ATen/native/transformers/hip/flash_attn/flash_common_hip.hpp>
+#include <ATen/native/transformers/hip/flash_attn/ck/me_ck_api.h>
+
+#include <iostream>
+
+namespace pytorch_flash {
+std::tuple<
+    at::Tensor, // output
+    at::Tensor, // q
+    at::Tensor, // k
+    at::Tensor, // v
+    at::Tensor, // lse
+    at::Tensor, // seed
+    at::Tensor, // offset
+    at::Tensor> // dropout randval
+mem_eff_forward_ck(
+  const at::Tensor& q,
+  const at::Tensor& k,
+  const at::Tensor& v,
+  float p_dropout,
+  bool return_dropout_randval,
+  std::optional<bool> is_causal,
+  std::optional<float> scale,
+  const std::optional<at::Tensor>& attn_bias_,
+  std::optional<at::Tensor>& out_,
+  const std::optional<at::Tensor>& cu_seqlens_q,
+  const std::optional<at::Tensor>& cu_seqlens_k,
+  const std::optional<at::Tensor>& seqstart_q,
+  const std::optional<at::Tensor>& seqstart_k,
+  std::optional<at::Generator> gen_,
+  std::optional<at::Tensor>& seqused_k_) {
+
+  std::cout << std::endl;
+  std::cout << "MADE IT INTO MY CODE " << std::endl;
+  // These normally get passed in as std::nullopt so just set to -1
+  // Note: See attention.cu ~line 928 and line 729
+  const int non_null_window_left  = -1;
+  const int non_null_window_right = -1;
+
+  TORCH_CHECK(
+    cu_seqlens_q.has_value() == cu_seqlens_k.has_value(),
+    "cu_seqlens_q and cu_seqlens_k must be both set or both not set");
+
+
+  // need to pass attn_bias to both of these
+  if(!seqstart_q.has_value()){
+    return mha_fwd_ck(
+      q,                     // q
+      k,                     // k
+      v,                     // v
+      out_,                  // opt(out_)
+      p_dropout,             // p_dropout
+      scale.value(),         // opt(softmax_scale)
+      is_causal.value(),     // opt(is_causal)
+      non_null_window_left,  // window_size_left
+      non_null_window_right, // window_size_right
+      false,                 // return_softmax/return_debug_mask
+      gen_,                  // gen
+      attn_bias_);           // attn_bias
+  } else {
+    // max sequence lengths are now at T.size(1) since q,k,v were all transposed
+    // in _scaled_dot_product_efficient_attention_cuda
+    const int64_t max_seqlen_q = q.size(1);
+    const int64_t max_seqlen_k = k.size(1);
+
+    return mha_varlen_fwd_ck(
+      q,                     // q
+      k,                     // k
+      v,                     // v
+      out_,                  // opt(out)
+      seqstart_q.value(),  // cu_seqlens_q
+      seqstart_k.value(),  // cu_seqlens_k
+      seqused_k_,            // opt(seqused_k)
+      max_seqlen_q,          // max_seqlen_q
+      max_seqlen_k,          // max_seqlen_k
+      p_dropout,             // p_dropout
+      scale.value(),         // softmax_scale
+      false,                 // zero_tensors
+      is_causal.value(),     // is_causal
+      non_null_window_left,  // window_size_left
+      non_null_window_right, // window_size_right
+      false,                 // return_softmax/return_debug_mask
+      gen_,                  // gen
+      attn_bias_);           // attn_bias
+  }
+}
+
+} // namespace pytorch_flash
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_bwd_ck.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_bwd_ck.hip
index 28bd893da0f34..a859c3bb11334 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_bwd_ck.hip
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_bwd_ck.hip
@@ -5,6 +5,7 @@
 #include <ATen/native/transformers/hip/flash_attn/flash_common_hip.hpp>
 #include <fmha_bwd.hpp>
 #include <mask.hpp>
+#include <iostream>
 
 namespace pytorch_flash {
 
@@ -12,16 +13,17 @@ fmha_bwd_traits get_ck_fmha_bwd_traits(const mask_info &mask,
                                        std::string dtype,
                                        int head_size,
                                        bool has_dropout,
-                                       bool enable_alibi,
-                                       bool deterministic)
+                                       bool enable_bias,
+                                       bool deterministic,
+                                       bool bias_requires_grad)
 {
     return fmha_bwd_traits{head_size,
                            head_size,
                            dtype,
                            false, // is_group_mode
                            mask.type,
-                           enable_alibi ? bias_enum::alibi : bias_enum::no_bias,
-                           false,    // has_dbias
+                           enable_bias ? bias_enum::elementwise_bias : bias_enum::no_bias,
+                           bias_requires_grad,    // has_dbias
                            has_dropout,
                            false, // s_randval
                            deterministic};
@@ -39,7 +41,9 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                                    const at::Tensor q,
                                    const at::Tensor k,
                                    const at::Tensor v,
-                                   std::optional<at::Tensor> &alibi_slopes_,
+                                   std::optional<at::Tensor> &attn_bias_,
+                                   bool bias_requires_grad,
+                                   std::optional<at::Tensor> &grad_bias,
                                    const at::Tensor out,
                                    const at::Tensor softmax_lse,
                                    const at::Tensor dout,
@@ -105,25 +109,48 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
     ck_tile::index_t stride_dq_acc = dq_acc.stride(2);
     ck_tile::index_t nhead_stride_dq_acc = dq_acc.stride(3);
 
+    //TODO_ANDY: need to add some stuff above for bias
+
+
+    // dbias: (batch_size, nheads, seqlen_q, seqlen_k) // TODO_ANDY verify this
+
     float p_undrop = 1.0 - p_dropout;
 
-    void *alibi_slopes_ptr = nullptr;
-    ck_tile::index_t stride_alibi_slopes = 0;
-
-    if (alibi_slopes_.has_value()) {
-        auto alibi_slopes = alibi_slopes_.value();
-        CHECK_DEVICE(alibi_slopes);
-        TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
-        TORCH_CHECK(alibi_slopes.sizes() == at::IntArrayRef({h}) || alibi_slopes.sizes() == at::IntArrayRef({b, h}));
-        alibi_slopes_ptr = alibi_slopes.data_ptr();
-        // alibi_slopes:(batch_size, nheads) or (nhead)
-        stride_alibi_slopes = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
+    void *attn_bias_ptr = nullptr;
+    ck_tile::index_t nhead_stride_bias = 0;
+    ck_tile::index_t batch_stride_bias = 0;
+    ck_tile::index_t stride_attn_bias = 0;
+
+    // bias: (batch_size, nheads, seqlen_q, seqlen_k)
+    if (attn_bias_.has_value()) {
+        auto a_b = attn_bias_.value();
+        CHECK_DEVICE(a_b);
+        TORCH_CHECK(a_b.stride(-1) == 1, "Attention bias tensor must have contiguous last dimension");
+        attn_bias_ptr = a_b.data_ptr();
+        stride_attn_bias = a_b.stride(2);
+        nhead_stride_bias = a_b.stride(1);
+        batch_stride_bias = a_b.stride(0);
+    }
+
+    void *dbias_ptr = nullptr;
+    ck_tile::index_t stride_dbias = 0;
+    ck_tile::index_t nhead_stride_dbias = 0;
+    ck_tile::index_t batch_stride_dbias = 0;
+    // dbias: (batch_size, nheads, seqlen_q, seqlen_k)
+    if(bias_requires_grad) {
+        // If bias_requires_grad is true, grad_bias is guaranteed to have a value via line 270
+        //grad_bias
+        auto dbias = grad_bias.value();
+        dbias_ptr = dbias.data_ptr();
+        stride_dbias = dbias.stride(2);
+        nhead_stride_dbias = dbias.stride(1);
+        batch_stride_dbias = dbias.stride(0);
     }
 
     return fmha_bwd_args{q.data_ptr(),
                          k.data_ptr(),
                          v.data_ptr(),
-                         alibi_slopes_ptr, // bias
+                         attn_bias_ptr, // bias
                          out.data_ptr(),
                          softmax_lse.data_ptr(),
                          dout.data_ptr(),
@@ -132,7 +159,7 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          dq.data_ptr(),
                          dk.data_ptr(),
                          dv.data_ptr(),
-                         nullptr, // dbias
+                         dbias_ptr, // dbias
                          dq_acc.data_ptr(), // dq_acc
                          nullptr, // seqstart_q
                          nullptr, // seqstart_k
@@ -150,7 +177,7 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          stride_q,
                          stride_k,
                          stride_v,
-                         stride_alibi_slopes,
+                         stride_attn_bias,
                          stride_o,
                          0, // stride_randval
                          stride_do,
@@ -158,11 +185,11 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          stride_dq,
                          stride_dk,
                          stride_dv,
-                         0, // stride_dbias, FA without bias
+                         stride_dbias, // stride_dbias, FA without bias TODO_ANDY: will probably need these
                          nhead_stride_q,
                          nhead_stride_k,
                          nhead_stride_v,
-                         0, // nhead_stride_bias, FA without bias
+                         nhead_stride_bias, // nhead_stride_bias, FA without bias
                          nhead_stride_o,
                          0, // nhead_stride_randval
                          nhead_stride_do,
@@ -171,11 +198,11 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          nhead_stride_dq,
                          nhead_stride_dk,
                          nhead_stride_dv,
-                         0, // nhead_stride_dbias, FA without dbias
+                         nhead_stride_dbias, // nhead_stride_dbias, FA without dbias
                          batch_stride_q,
                          batch_stride_k,
                          batch_stride_v,
-                         0  , // batch_stride_bias, FA without bias
+                         batch_stride_bias, // batch_stride_bias, FA without bias
                          batch_stride_o,
                          0, // batch_stride_randval
                          batch_stride_do,
@@ -184,7 +211,7 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          batch_stride_dq,
                          batch_stride_dk,
                          batch_stride_dv,
-                         0  , // batch_stride_dbias, FA without dbias
+                         batch_stride_dbias, // batch_stride_dbias, FA without dbias
                          split_stride_dq_acc,
                          mask.left,
                          mask.right,
@@ -193,8 +220,8 @@ fmha_bwd_args get_ck_fmha_bwd_args(const mask_info &mask,
                          p_undrop,
                          drop_seed_offset};
 }
-
-std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor>
+//START HERE ANDY - JUST ADDED THE FIFTH RETURN TYPE, MAKE SURE WE ARE RETURNING DBIAS
+std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
 mha_bwd_ck(const at::Tensor &dout,                   // batch_size x seqlen_q x num_heads, x head_size_og
            const at::Tensor &q,                      // batch_size x seqlen_q x num_heads x head_size
            const at::Tensor &k,                      // batch_size x seqlen_k x num_heads_k x head_size
@@ -204,7 +231,9 @@ mha_bwd_ck(const at::Tensor &dout,                   // batch_size x seqlen_q x
            std::optional<at::Tensor> &dq_,           // batch_size x seqlen_q x num_heads x head_size
            std::optional<at::Tensor> &dk_,           // batch_size x seqlen_k x num_heads_k x head_size
            std::optional<at::Tensor> &dv_,           // batch_size x seqlen_k x num_heads_k x head_size
-           std::optional<at::Tensor> &alibi_slopes_, // num_heads or batch_size x num_heads
+           std::optional<at::Tensor> &attn_bias_, // num_heads or batch_size x num_heads
+           bool bias_requires_grad,
+           std::optional<at::Tensor> &grad_bias,
            const float p_dropout,                    // probability to drop
            const float softmax_scale,
            const bool is_causal,
@@ -242,6 +271,9 @@ mha_bwd_ck(const at::Tensor &dout,                   // batch_size x seqlen_q x
     TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
     TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
 
+    TORCH_CHECK((bias_requires_grad && grad_bias.has_value()) || (!bias_requires_grad),
+            "If bias_requires_grad is set, grad_bias must have a value");
+
     const auto sizes = q.sizes();
 
     const int batch_size = sizes[0];
@@ -354,7 +386,13 @@ mha_bwd_ck(const at::Tensor &dout,                   // batch_size x seqlen_q x
         ck_tile::stream_config stream_config{stream};
         dq.zero_(); // ck use atomic operation on dq
         auto traits =
-            get_ck_fmha_bwd_traits(mask, q_dtype_str, head_size_8x, is_dropout, alibi_slopes_.has_value(), deterministic);
+            get_ck_fmha_bwd_traits(mask,
+                                   q_dtype_str,
+                                   head_size_8x,
+                                   is_dropout,
+                                   attn_bias_.has_value(),
+                                   deterministic,
+                                   bias_requires_grad);
 
         auto args =
             get_ck_fmha_bwd_args(
@@ -368,7 +406,9 @@ mha_bwd_ck(const at::Tensor &dout,                   // batch_size x seqlen_q x
                 q,
                 k,
                 v,
-                alibi_slopes_,
+                attn_bias_,
+                bias_requires_grad,
+                grad_bias,
                 out,
                 softmax_lse,
                 dout_padded,
@@ -400,6 +440,15 @@ mha_bwd_ck(const at::Tensor &dout,                   // batch_size x seqlen_q x
         dv = dv.index({"...", at::indexing::Slice(at::indexing::None, head_size_og)});
     }
 
-    return { dq, dk, dv, softmax_d };
+    //TODO_ANDY need to return dGrad also
+    at::Tensor dbias;
+    if(bias_requires_grad) {
+        dbias = grad_bias.value();
+    } else {
+        dbias = at::empty({batch_size, num_heads, seqlen_q, seqlen_k}, q.options());
+    }
+
+    std::cout << "MHA_BWD_CK RAN AND COMPLETED" << std::endl;
+    return { dq, dk, dv, softmax_d, dbias };
 }
 } // namespace pytorch_flash
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_fwd_ck.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_fwd_ck.hip
index 4d7726cec8fd3..236c1df2f4470 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_fwd_ck.hip
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_fwd_ck.hip
@@ -6,7 +6,7 @@
 #include <fmha_fwd.hpp>
 #include <mask.hpp>
 
-
+#include <iostream>
 
 namespace pytorch_flash {
 
@@ -16,7 +16,7 @@ fmha_fwd_traits get_ck_fmha_fwd_traits(const mask_info &mask,
                                        int head_size,
                                        bool has_dropout,
                                        bool has_lse,
-                                       bool enable_alibi)
+                                       bool enable_bias)
 {
     return fmha_fwd_traits{head_size,
                            head_size,
@@ -24,7 +24,7 @@ fmha_fwd_traits get_ck_fmha_fwd_traits(const mask_info &mask,
                            false, // is_group_mode
                            true,  // is_v_rowmajor
                            mask.type,
-                           enable_alibi ? bias_enum::alibi : bias_enum::no_bias,
+                           enable_bias ? bias_enum::elementwise_bias : bias_enum::no_bias,
                            has_lse,
                            has_dropout,
                            false}; // do_fp8_static_quant
@@ -44,7 +44,7 @@ fmha_fwd_args get_ck_fmha_fwd_args(bool has_lse,
                                    const at::Tensor q,
                                    const at::Tensor k,
                                    const at::Tensor v,
-                                   std::optional<at::Tensor> &alibi_slopes_,
+                                   std::optional<at::Tensor> &attn_bias_,
                                    at::Tensor out,
                                    at::Tensor softmax_lse,
                                    at::Tensor dropout_randval,
@@ -57,7 +57,7 @@ fmha_fwd_args get_ck_fmha_fwd_args(bool has_lse,
     // v: (batch_size, seqlen_k, nheads_k, d)
     // o: (batch_size, seqlen_q, nheads, d)
 
-    // alibi_slopes:(batch_size, nheads) or (nhead)
+    // attn_bias: (batch_size, nheads, seqlen_q, seqlen_k)
     // lse: (batch_size, nheads, seqlen_q)
     // randval: (batch_size, nheads, seqlen_q, seqlen_k)
 
@@ -82,22 +82,30 @@ fmha_fwd_args get_ck_fmha_fwd_args(bool has_lse,
     ck_tile::index_t batch_stride_lse = has_lse ? softmax_lse.stride(0) : 0;
     ck_tile::index_t batch_stride_randval = has_dropout_randval ? dropout_randval.stride(0) : 0;
 
-    void *alibi_slopes_ptr = nullptr;
-    ck_tile::index_t stride_alibi_slopes = 0;
+    void *attn_bias_ptr = nullptr;
+    ck_tile::index_t stride_attn_bias = 0;
+
+    if (attn_bias_.has_value()) {
+        auto a_b = attn_bias_.value();
+        CHECK_DEVICE(a_b);
+        TORCH_CHECK(a_b.stride(-1) == 1, "attention bias tensor must have contiguous last dimension");
+        // Following check was a remnant of the expectation that the bias tensor would be alibi which only has 2 dimensions
+        // whereas what pytorch provides (elementwise bias) is a 4 dimensional tensor of the shape {b, h, s_q, s_k}
+        //TORCH_CHECK(alibi_slopes.sizes() == at::IntArrayRef({h}) || alibi_slopes.sizes() == at::IntArrayRef({b, h}));
+        attn_bias_ptr = a_b.data_ptr();
+        // Previously, looks like alibi slopes could be 1 of 2 shapes. either {h} or {b, h} so this check makes sure
+        // to grab the batch size if there are 2 dimensions, and a stride of zero if there's only one dimension.
+        // so assuming we are getting what pytorch wants to give us which is the aforementioned 4-d tensor, just grab the batch
+        // dimension
+        //stride_attn_bias = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
+        stride_attn_bias = a_b.stride(0);
 
-    if (alibi_slopes_.has_value()) {
-        auto alibi_slopes = alibi_slopes_.value();
-        CHECK_DEVICE(alibi_slopes);
-        TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
-        TORCH_CHECK(alibi_slopes.sizes() == at::IntArrayRef({h}) || alibi_slopes.sizes() == at::IntArrayRef({b, h}));
-        alibi_slopes_ptr = alibi_slopes.data_ptr();
-        stride_alibi_slopes = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
     }
 
     return fmha_fwd_args{q.data_ptr(),
                          k.data_ptr(),
                          v.data_ptr(),
-                         alibi_slopes_ptr, // bias
+                         attn_bias_ptr, // bias
                          has_dropout_randval ? dropout_randval.data_ptr() : nullptr,
                          has_lse ? softmax_lse.data_ptr() : nullptr,
                          out.data_ptr(),
@@ -118,13 +126,13 @@ fmha_fwd_args get_ck_fmha_fwd_args(bool has_lse,
                          stride_q,
                          stride_k,
                          stride_v,
-                         stride_alibi_slopes,
+                         stride_attn_bias,
                          stride_randval,
                          stride_o,
                          nhead_stride_q,
                          nhead_stride_k,
                          nhead_stride_v,
-                         0, // nhead_stride_bias, FA without bias
+                         0, // nhead_stride_bias, FA without bias : TODO_ANDY CHECK IF WE NEED TO DO SOMETHING WITH THIS
                          nhead_stride_randval,
                          nhead_stride_lse,
                          nhead_stride_o,
@@ -148,14 +156,14 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
            const at::Tensor &k,                      // batch_size x seqlen_k x num_heads_k x head_size
            const at::Tensor &v,                      // batch_size x seqlen_k x num_heads_k x head_size
            std::optional<at::Tensor> &out_,          // batch_size x seqlen_q x num_heads xhead_size
-           std::optional<at::Tensor> &alibi_slopes_, // num_heads or batch_size x num_heads
            const float p_dropout,
            const float softmax_scale,
            bool is_causal,
            int window_size_left,
            int window_size_right,
            const bool return_dropout_randval,
-           std::optional<at::Generator> gen_)
+           std::optional<at::Generator> gen_,
+           const std::optional<at::Tensor>& attn_bias_) // batch_size x nheads x seqlen_q x seqlen_k
 {
     auto q_dtype = q.dtype();
     TORCH_CHECK(q_dtype == at::kHalf || q_dtype == at::kBFloat16,
@@ -189,7 +197,7 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
     if (window_size_right >= seqlen_k) { window_size_right = -1; }
 
     // causal=true is the same as causal=false in this case
-    if (seqlen_q == 1 && !alibi_slopes_.has_value()) { is_causal = false; }
+    if (seqlen_q == 1 && !attn_bias_.has_value()) { is_causal = false; }
 
     mask_info mask;
     if (is_causal) {
@@ -209,7 +217,7 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
 
     // Faster to transpose q from (b, 1, (nheads_kv ngroups), d) to (b, ngroups, nheads_kv, d) in this case
     // H/t Daniel Haziza
-    const int seqlenq_ngroups_swapped = seqlen_q == 1 && num_heads > num_heads_k && window_size_left < 0 && window_size_right < 0 && p_dropout == 0.f && head_size % 8 == 0 && !alibi_slopes_.has_value();
+    const int seqlenq_ngroups_swapped = seqlen_q == 1 && num_heads > num_heads_k && window_size_left < 0 && window_size_right < 0 && p_dropout == 0.f && head_size % 8 == 0 && !attn_bias_.has_value();
     const int ngroups = num_heads / num_heads_k;
     at::Tensor temp_q = q;
     if (seqlenq_ngroups_swapped) {
@@ -217,7 +225,8 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
         seqlen_q = ngroups;
         num_heads = num_heads_k;
     }
-
+    std::cout << "MHA_FWD_CK: CHECKING temp_Q SHAPE: " << temp_q.sizes() << std::endl;
+    std::cout << "MHA_FWD_CK: CHECKING Q SHAPE     : " << q.sizes() << std::endl;
     CHECK_SHAPE(temp_q, batch_size, seqlen_q, num_heads, head_size);
     CHECK_SHAPE(k, batch_size, seqlen_k, num_heads_k, head_size);
     CHECK_SHAPE(v, batch_size, seqlen_k, num_heads_k, head_size);
@@ -305,6 +314,15 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
         offset_t = at::empty({}, at::dtype(at::kLong).device(at::kCUDA));
     }
 
+    // remove const from attn_bias_
+    // TODO: sanity check this
+    std::optional<at::Tensor> attn_bias;
+    if( attn_bias_.has_value())
+    {
+      std::cout << "CONFIRMED YOUR CODE IS GETTING HIT AND ATTENTION BIAS IS SET" << std::endl;
+      attn_bias = attn_bias_;
+    }
+
     if (seqlen_k > 0) {
         auto drop_seed_offset = std::make_pair(rng_state_ptr, rng_state_ptr + 1);
         auto stream = at::cuda::getCurrentHIPStream().stream();
@@ -317,7 +335,7 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
                 head_size_8x,
                 has_dropout,
                 has_lse,
-                alibi_slopes_.has_value());
+                attn_bias_.has_value());
 
         auto args =
             get_ck_fmha_fwd_args(
@@ -333,7 +351,7 @@ mha_fwd_ck(const at::Tensor &q,                      // batch_size x seqlen_q x
                 q,
                 k,
                 v,
-                alibi_slopes_,
+                attn_bias,
                 out,
                 softmax_lse,
                 p,
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_bwd_ck.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_bwd_ck.hip
index d57ead48e2b50..91f11e697cae5 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_bwd_ck.hip
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_bwd_ck.hip
@@ -14,16 +14,17 @@ fmha_bwd_traits get_ck_fmha_varlen_bwd_traits(const mask_info &mask,
                                               std::string dtype,
                                               int head_size,
                                               bool has_dropout,
-                                              bool enable_alibi,
-                                              bool deterministic)
+                                              bool enable_bias,
+                                              bool deterministic,
+                                              bool bias_requires_grad)
 {
     return fmha_bwd_traits{head_size,
                            head_size,
                            dtype,
                            true, // is_group_mode
                            mask.type,
-                           enable_alibi ? bias_enum::alibi : bias_enum::no_bias,
-                           false,    // has_dbias
+                           enable_bias ? bias_enum::elementwise_bias : bias_enum::no_bias,
+                           bias_requires_grad,    // has_dbias
                            has_dropout,
                            false, // s_randval
                            deterministic};
@@ -43,7 +44,9 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                                           const at::Tensor v,
                                           const at::Tensor seqlens_q,
                                           const at::Tensor seqlens_k,
-                                          std::optional<at::Tensor> &alibi_slopes_,
+                                          std::optional<at::Tensor> &attn_bias_,
+                                          bool bias_requires_grad,
+                                          std::optional<at::Tensor> &grad_bias,
                                           const at::Tensor out,
                                           const at::Tensor softmax_lse,
                                           const at::Tensor dout,
@@ -115,23 +118,41 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
 
     float p_undrop = 1.0 - p_dropout;
 
-    void *alibi_slopes_ptr = nullptr;
-    ck_tile::index_t stride_alibi_slopes = 0;
-
-    if (alibi_slopes_.has_value()) {
-        auto alibi_slopes = alibi_slopes_.value();
-        CHECK_DEVICE(alibi_slopes);
-        TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
-        TORCH_CHECK(alibi_slopes.sizes() == at::IntArrayRef({h}) || alibi_slopes.sizes() == at::IntArrayRef({b, h}));
-        alibi_slopes_ptr = alibi_slopes.data_ptr();
-        // alibi_slopes:(batch_size, nheads) or (nhead)
-        stride_alibi_slopes = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
+    //TODO_ANDY: Probably need to handle some bias stuff similar to the above
+    // bias: (batch_size, nheads, seqlen_q, seqlen_k)
+    void *attn_bias_ptr = nullptr;
+    ck_tile::index_t nhead_stride_bias = 0;
+    ck_tile::index_t batch_stride_bias = 0;
+    ck_tile::index_t stride_attn_bias = 0;
+    if (attn_bias_.has_value()) {
+        auto a_b = attn_bias_.value();
+        CHECK_DEVICE(a_b);
+        TORCH_CHECK(a_b.stride(-1) == 1, "Attention bias tensor must have contiguous last dimension");
+        attn_bias_ptr = a_b.data_ptr();
+        stride_attn_bias = a_b.stride(2);
+        nhead_stride_bias = a_b.stride(1);
+        batch_stride_bias = a_b.stride(0);
+    }
+
+    void *dbias_ptr = nullptr;
+    ck_tile::index_t stride_dbias = 0;
+    ck_tile::index_t nhead_stride_dbias = 0;
+    ck_tile::index_t batch_stride_dbias = 0;
+    // dbias: (batch_size, nheads, seqlen_q, seqlen_k)
+    if(bias_requires_grad) {
+        // If bias_requires_grad is true, grad_bias is guaranteed to have a value via line 270
+        //grad_bias
+        auto dbias = grad_bias.value();
+        dbias_ptr = dbias.data_ptr();
+        stride_dbias = dbias.stride(2);
+        nhead_stride_dbias = dbias.stride(1);
+        batch_stride_dbias = dbias.stride(0);
     }
 
     return fmha_bwd_args{q.data_ptr(),
                          k.data_ptr(),
                          v.data_ptr(),
-                         alibi_slopes_ptr, // bias
+                         attn_bias_ptr, // bias
                          out.data_ptr(),
                          softmax_lse.data_ptr(),
                          dout.data_ptr(),
@@ -140,7 +161,7 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          dq.data_ptr(),
                          dk.data_ptr(),
                          dv.data_ptr(),
-                         nullptr, // dbias
+                         dbias_ptr, // dbias
                          dq_acc.data_ptr(), // dq_acc
                          seqlens_q.data_ptr(), // seqstart_q
                          seqlens_k.data_ptr(), // seqstart_k
@@ -158,7 +179,7 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          stride_q,
                          stride_k,
                          stride_v,
-                         stride_alibi_slopes,
+                         stride_attn_bias,
                          stride_o,
                          0, // stride_randval
                          stride_do,
@@ -166,11 +187,11 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          stride_dq,
                          stride_dk,
                          stride_dv,
-                         0, // stride_dbias, FA without bias
+                         stride_dbias, // stride_dbias, FA without bias
                          nhead_stride_q,
                          nhead_stride_k,
                          nhead_stride_v,
-                         0, // nhead_stride_bias, FA without bias
+                         nhead_stride_bias, // nhead_stride_bias, FA without bias
                          nhead_stride_o,
                          0, // nhead_stride_randval
                          nhead_stride_do,
@@ -179,11 +200,11 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          nhead_stride_dq,
                          nhead_stride_dk,
                          nhead_stride_dv,
-                         0, // nhead_stride_dbias, FA without dbias
+                         nhead_stride_dbias, // nhead_stride_dbias, FA without dbias
                          batch_stride_q,
                          batch_stride_k,
                          batch_stride_v,
-                         0  , // batch_stride_bias, FA without bias
+                         batch_stride_bias, // batch_stride_bias, FA without bias
                          batch_stride_o,
                          0, // batch_stride_randval
                          batch_stride_do,
@@ -192,7 +213,7 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          batch_stride_dq,
                          batch_stride_dk,
                          batch_stride_dv,
-                         0  , // batch_stride_dbias, FA without dbias
+                         batch_stride_dbias, // batch_stride_dbias, FA without dbias
                          split_stride_dq_acc,
                          mask.left,
                          mask.right,
@@ -202,7 +223,7 @@ fmha_bwd_args get_ck_fmha_varlen_bwd_args(const mask_info &mask,
                          drop_seed_offset};
 }
 
-std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor>
+std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
 mha_varlen_bwd_ck(const at::Tensor &dout,                   // total_q x num_heads x head_size
                   const at::Tensor &q,                      // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
                   const at::Tensor &k,                      // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
@@ -214,7 +235,9 @@ mha_varlen_bwd_ck(const at::Tensor &dout,                   // total_q x num_hea
                   std::optional<at::Tensor> &dv_,           // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
                   const at::Tensor &cu_seqlens_q,           // b+1
                   const at::Tensor &cu_seqlens_k,           // b+1
-                  std::optional<at::Tensor> &alibi_slopes_, // num_heads or b x num_heads
+                  std::optional<at::Tensor> &attn_bias_,    // b x num_heads x seqlen_q x seqlen_k
+                  bool bias_requires_grad,
+                  std::optional<at::Tensor> &grad_bias,
                   const int max_seqlen_q,
                   const int max_seqlen_k, // max sequence length to choose the kernel
                   const float p_dropout,  // probability to drop
@@ -260,6 +283,9 @@ mha_varlen_bwd_ck(const at::Tensor &dout,                   // total_q x num_hea
     CHECK_CONTIGUOUS(cu_seqlens_q);
     CHECK_CONTIGUOUS(cu_seqlens_k);
 
+    TORCH_CHECK((bias_requires_grad && grad_bias.has_value()) || (!bias_requires_grad),
+            "If bias_requires_grad is set, grad_bias must have a value");
+
     const auto sizes = q.sizes();
 
     const int total_q = sizes[0];
@@ -381,7 +407,13 @@ mha_varlen_bwd_ck(const at::Tensor &dout,                   // total_q x num_hea
         ck_tile::stream_config stream_config{stream};
         dq.zero_(); // ck use atomic operation on dq
         auto traits =
-            get_ck_fmha_varlen_bwd_traits(mask, q_dtype_str, head_size_8x, is_dropout, alibi_slopes_.has_value(), deterministic);
+            get_ck_fmha_varlen_bwd_traits(mask,
+                                          q_dtype_str,
+                                          head_size_8x,
+                                          is_dropout,
+                                          attn_bias_.has_value(),
+                                          deterministic,
+                                          bias_requires_grad);
 
         auto args =
             get_ck_fmha_varlen_bwd_args(
@@ -397,7 +429,9 @@ mha_varlen_bwd_ck(const at::Tensor &dout,                   // total_q x num_hea
                 v,
                 cu_seqlens_q,
                 cu_seqlens_k,
-                alibi_slopes_,
+                attn_bias_,
+                bias_requires_grad,
+                grad_bias,
                 out,
                 softmax_lse,
                 dout_padded,
@@ -428,7 +462,15 @@ mha_varlen_bwd_ck(const at::Tensor &dout,                   // total_q x num_hea
         dk = dk.index({"...", at::indexing::Slice(at::indexing::None, head_size_og)});
         dv = dv.index({"...", at::indexing::Slice(at::indexing::None, head_size_og)});
     }
+    at::Tensor dbias;
+    if(bias_requires_grad) {
+        dbias = grad_bias.value();
+    } else {
+        dbias = at::empty({batch_size, num_heads, max_seqlen_q, max_seqlen_k}, q.options());
+    }
+
 
-    return { dq, dk, dv, softmax_d };
+    return { dq, dk, dv, softmax_d, dbias };
+    std::cout << "TOUCHING VARLEN BWD" << std::endl;
 }
 } // namespace pytorch_flash
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_fwd_ck.hip b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_fwd_ck.hip
index 2ef2ab24d9a60..7d3ed60756200 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_fwd_ck.hip
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/ck/mha_varlen_fwd_ck.hip
@@ -13,7 +13,7 @@ fmha_fwd_traits get_ck_fmha_varlen_fwd_traits(const mask_info &mask,
                                               int head_size,
                                               bool has_dropout,
                                               bool has_lse,
-                                              bool enable_alibi)
+                                              bool enable_bias)
 {
     return fmha_fwd_traits{head_size,
                            head_size,
@@ -21,7 +21,7 @@ fmha_fwd_traits get_ck_fmha_varlen_fwd_traits(const mask_info &mask,
                            true, // is_group_mode
                            true, // is_v_rowmajor
                            mask.type,
-                           enable_alibi ? bias_enum::alibi : bias_enum::no_bias,
+                           enable_bias ? bias_enum::elementwise_bias : bias_enum::no_bias,
                            has_lse,
                            has_dropout,
                            false}; // do_fp8_static_quant
@@ -42,11 +42,10 @@ fmha_fwd_args get_ck_fmha_varlen_fwd_args(bool has_lse,
                                           const at::Tensor v,
                                           const at::Tensor seqlens_q,
                                           const at::Tensor seqlens_k,
-                                          std::optional<at::Tensor> &alibi_slopes_,
+                                          std::optional<at::Tensor> &attn_bias_,
                                           at::Tensor out,
                                           at::Tensor softmax_lse,
                                           at::Tensor dropout_randval,
-
                                           float softmax_scale,
                                           float p_dropout,
                                           std::pair<uint64_t*, uint64_t*> drop_seed_offset)
@@ -56,7 +55,7 @@ fmha_fwd_args get_ck_fmha_varlen_fwd_args(bool has_lse,
     // v: (total_k, nheads_k, d)
     // o: (total_q, nheads, d)
 
-    // alibi_slopes:(batch, nheads) or (nhead)
+    // attn_bias :(batch, nheads, max_seqlen_q, max_seqlen_k)
     // lse: (batch, nheads, max_seqlen_q)
     // randval: (nheads, total_q, max_seqlen_k)
 
@@ -84,22 +83,23 @@ fmha_fwd_args get_ck_fmha_varlen_fwd_args(bool has_lse,
     ck_tile::index_t batch_stride_lse = has_lse ? softmax_lse.stride(0) : 0;
     ck_tile::index_t batch_stride_randval = 0;
 
-    void *alibi_slopes_ptr = nullptr;
-    ck_tile::index_t stride_alibi_slopes = 0;
-
-    if (alibi_slopes_.has_value()) {
-        auto alibi_slopes = alibi_slopes_.value();
-        CHECK_DEVICE(alibi_slopes);
-        TORCH_CHECK(alibi_slopes.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
-        TORCH_CHECK(alibi_slopes.sizes() == at::IntArrayRef({h}) || alibi_slopes.sizes() == at::IntArrayRef({b, h}));
-        alibi_slopes_ptr = alibi_slopes.data_ptr();
-        stride_alibi_slopes = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
+    void *attn_bias_ptr = nullptr;
+    ck_tile::index_t stride_attn_bias = 0;
+
+    if (attn_bias_.has_value()) {
+        auto a_b = attn_bias_.value();
+        CHECK_DEVICE(a_b);
+        TORCH_CHECK(a_b.stride(-1) == 1, "ALiBi slopes tensor must have contiguous last dimension");
+        //TORCH_CHECK(alibi_slopes.sizes() == at::IntArrayRef({h}) || alibi_slopes.sizes() == at::IntArrayRef({b, h}));
+        attn_bias_ptr = a_b.data_ptr();
+        //stride_attn_bias = alibi_slopes.dim() == 2 ? alibi_slopes.stride(0) : 0;
+        stride_attn_bias = a_b.stride(0);
     }
 
     return fmha_fwd_args{q.data_ptr(),
                          k.data_ptr(),
                          v.data_ptr(),
-                         alibi_slopes_ptr, // bias
+                         attn_bias_ptr, // bias
                          has_dropout_randval ? dropout_randval.data_ptr() : nullptr,
                          has_lse ? softmax_lse.data_ptr() : nullptr,
                          out.data_ptr(),
@@ -120,7 +120,7 @@ fmha_fwd_args get_ck_fmha_varlen_fwd_args(bool has_lse,
                          stride_q,
                          stride_k,
                          stride_v,
-                         stride_alibi_slopes,
+                         stride_attn_bias,
                          stride_randval,
                          stride_o,
                          nhead_stride_q,
@@ -153,7 +153,6 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
                   const at::Tensor &cu_seqlens_q,  // b+1
                   const at::Tensor &cu_seqlens_k,  // b+1
                   std::optional<at::Tensor> & /*seqused_k*/,
-                  std::optional<at::Tensor> &alibi_slopes_, // num_heads or b x num_heads
                   int max_seqlen_q,
                   const int max_seqlen_k,
                   const float p_dropout,
@@ -163,8 +162,10 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
                   int window_size_left,
                   int window_size_right,
                   const bool return_dropout_randval,
-                  std::optional<at::Generator> gen_)
+                  std::optional<at::Generator> gen_,
+                  const std::optional<at::Tensor>& attn_bias_)
 {
+    std::cout << "MHA_VARLEN_BWD_CK ENTER" << std::endl;
     auto q_dtype = q.dtype();
     TORCH_CHECK(q_dtype == at::kHalf || q_dtype == at::kBFloat16,
                 "FlashAttention only support fp16 and bf16 data type");
@@ -200,7 +201,7 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
     const int max_num_blocks_per_seq = 0;
     const int num_blocks = 0;
 
-    if (max_seqlen_q == 1 && !alibi_slopes_.has_value()) { is_causal = false; }  // causal=true is the same as causal=false in this case
+    if (max_seqlen_q == 1 && !attn_bias_.has_value()) { is_causal = false; }  // causal=true is the same as causal=false in this case
 
     // TODO
     // Faster to transpose q from (b, 1, (nheads_kv ngroups), d) to (b, ngroups, nheads_kv, d) in this case
@@ -233,7 +234,24 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
         mask = mask_info::decode(mask_identify, max_seqlen_q, max_seqlen_k); // local
     }
 
-    CHECK_SHAPE(q, total_q, num_heads, head_size_og);
+
+    std::cout << "ABOUT TO CHECK Q's shape: " << q.sizes() << std::endl;
+    std::cout << "total_q     : " << total_q << std::endl;
+    std::cout << "num_heads   : " << num_heads << std::endl;
+    std::cout << "head_size_og: " << head_size_og << std::endl;
+
+    std::cout << "ABOUT TO CHECK K's shape: " << k.sizes() << std::endl;
+    std::cout << "total_k     : " << total_k << std::endl;
+    std::cout << "num_heads_k : " << num_heads_k << std::endl;
+    std::cout << "head_size_og: " << head_size_og << std::endl;
+
+    std::cout << "ABOUT TO CHECK V's shape: " << v.sizes() << std::endl;
+    std::cout << "total_v     : " << total_k << std::endl;
+    std::cout << "num_heads_v : " << num_heads_k << std::endl;
+    std::cout << "head_size_og: " << head_size_og << std::endl;
+
+
+    //CHECK_SHAPE(q, total_q, num_heads, head_size_og);
     CHECK_SHAPE(k, total_k, num_heads_k, head_size_og);
     CHECK_SHAPE(v, total_k, num_heads_k, head_size_og);
     CHECK_SHAPE(cu_seqlens_q, batch_size + 1);
@@ -307,6 +325,13 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
             flash::ParsePhiloxCudaState, dim3(1), dim3(64), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), philox_args, rng_state_ptr);
     }
 
+    // remove const from attn_bias_
+    std::optional<at::Tensor> attn_bias;
+    if( attn_bias_.has_value())
+    {
+      attn_bias = attn_bias_;
+    }
+
 
     if (max_seqlen_k > 0) {
         auto drop_seed_offset = std::make_pair(rng_state_ptr, rng_state_ptr + 1);
@@ -314,7 +339,7 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
         ck_tile::stream_config stream_config{stream};
 
         auto traits =
-            get_ck_fmha_varlen_fwd_traits(mask, q_dtype_str, head_size_8x, has_dropout, has_lse, alibi_slopes_.has_value());
+            get_ck_fmha_varlen_fwd_traits(mask, q_dtype_str, head_size_8x, has_dropout, has_lse, attn_bias_.has_value());
 
         auto args =
             get_ck_fmha_varlen_fwd_args(
@@ -331,7 +356,7 @@ mha_varlen_fwd_ck(const at::Tensor &q,                   // total_q x num_heads
                 v_padded,
                 cu_seqlens_q,
                 cu_seqlens_k,
-                alibi_slopes_,
+                attn_bias,
                 out,
                 softmax_lse,
                 p,
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/rename_ck_autogen_files.output.txt b/aten/src/ATen/native/transformers/hip/flash_attn/ck/rename_ck_autogen_files.output.txt
deleted file mode 100644
index 78f844fd2a1e0..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/rename_ck_autogen_files.output.txt
+++ /dev/null
@@ -1,1810 +0,0 @@
-fmha_bwd_api.hip -> fmha_ck_autogen_5919133d2ed892745013b2fc5d503414cf0a4d83.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2.hip -> fmha_ck_autogen_e11a3b7d4fdfed64e64f7a95dbc64eff541092d6.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_deterministic.hip -> fmha_ck_autogen_01cb354dddef6e99e4ac843f2adafcddfc58d520.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_pd.hip -> fmha_ck_autogen_1b3e7c8969027d3316875f33dc50fe022e05ce37.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_38273a2f8e6bbb42ba0b0871b6c95abb34531f33.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_ps.hip -> fmha_ck_autogen_2d43460c011b8d5e01ea98c9b8ddce962de59a96.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_4c0c50a1fac82d47dff2357ee3ddbfa0b2c8d487.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_psd.hip -> fmha_ck_autogen_2a3a980a26682d879c3a3425f3ba5be3f5761adf.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_008f2429c678d13386a06e8d8b15c4b480940ff3.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_group_o2_ps.hip -> fmha_ck_autogen_811db756577b61cde9fe8279d956980db9ee21a4.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_group_o2_ps_deterministic.hip -> fmha_ck_autogen_492fbc418e829f89bcb8d93f8afd2869dd8dfccc.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_group_o2_psd.hip -> fmha_ck_autogen_75f2010bf6c478d2f0eba77e912697661306c1cb.hip
-fmha_bwd_convert_dq_d128_bf16_b64x128_group_o2_psd_deterministic.hip -> fmha_ck_autogen_0153ec18d3ded0f8bdc6459ea5757ebd94d9faf2.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2.hip -> fmha_ck_autogen_3eb2ea922daabbba131b90713e06d8caf5f30662.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_deterministic.hip -> fmha_ck_autogen_c0f76aff077c28f8afd7b22f284cf2894e08a043.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_pd.hip -> fmha_ck_autogen_f48f8b681a405bfeba5aadaef40f32367ec5cd2b.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_4cabdafad0bf803223ba5e8f474cd59233dc48cb.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_ps.hip -> fmha_ck_autogen_0801c56831b4c6428200db6318638a2129bb197a.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_91b9e2616c2fe0480096b1ccf0f74d584b220146.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_psd.hip -> fmha_ck_autogen_4f1e1c969b57659e7e1367ac9ba10ed5ef5b69a9.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_ecd7dec90b3c62bf3a30bd75d3c6869529a06b01.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_group_o2_ps.hip -> fmha_ck_autogen_88ea5b5346c87cc4fc1e841c518080df4ab811a2.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_group_o2_ps_deterministic.hip -> fmha_ck_autogen_4395d3c96b3f4556b9765fd0a3b5701b2fb10948.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_group_o2_psd.hip -> fmha_ck_autogen_b8fbc6f6e9c515edce3c7a438b3bc308b30d3857.hip
-fmha_bwd_convert_dq_d128_fp16_b64x128_group_o2_psd_deterministic.hip -> fmha_ck_autogen_490a68220a7b621ae9817d7b77f55de239b0a4f3.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2.hip -> fmha_ck_autogen_344932e2655d7b32704be8de9a63bbd8c3369f02.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_deterministic.hip -> fmha_ck_autogen_5a85ae0a16e4b293b549bcb6a3ee52df7fccca32.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_pd.hip -> fmha_ck_autogen_963986150adcd6e1d3886bacf2166de1252e14df.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_8bd1a40b12ce927323594fcce61eb9c20cc5e3d4.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_ps.hip -> fmha_ck_autogen_296c5836ba118969c4ba89ed62a98dffe3105738.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_6cfb7075345704340ff33dc0ef7c04ef127f26ad.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_psd.hip -> fmha_ck_autogen_22511de2592b6e350737e44865e1fed6496e3f32.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_609f68180582384ba81aae2b1d4a4c52dde2c68c.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_group_o2_ps.hip -> fmha_ck_autogen_c9fe51f982abd60e567d4238d3266fb60e45814b.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_group_o2_ps_deterministic.hip -> fmha_ck_autogen_10a055e5c3d6a953d470db5dc21449766248058a.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_group_o2_psd.hip -> fmha_ck_autogen_327e27892bc57f3dec0da24f94f2a483d6c9321b.hip
-fmha_bwd_convert_dq_d256_bf16_b64x64_group_o2_psd_deterministic.hip -> fmha_ck_autogen_c581974c8b6f43f60d0af29c350d850b55c03121.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2.hip -> fmha_ck_autogen_01ac1a2ecf9a487809e46faa92e267df2d47de91.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_deterministic.hip -> fmha_ck_autogen_dbc4135fce01e8731fec7a78d0cc0fdeeae28b90.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_pd.hip -> fmha_ck_autogen_e09d9baa269dfbb30b714389d1733be51cc419b7.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_5f71e663978dbcba859c5114ec675a712e343fd6.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_ps.hip -> fmha_ck_autogen_d257148f457557ea80ca56690e525db3a4b0ff55.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_8e2c587db8bd9f1b551624e0cf8b67a90245d7da.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_psd.hip -> fmha_ck_autogen_8c13c4f3f645a2bb475eb1c55ce1de452f0e2332.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_7b7fa76609243a8709f349ffc0d9d88157f28dc9.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_group_o2_ps.hip -> fmha_ck_autogen_2b3326e055da32cc979892a2fbd0f7b003cb9f98.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_group_o2_ps_deterministic.hip -> fmha_ck_autogen_671828f15eec2a58be23063a1a8132d337cd26de.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_group_o2_psd.hip -> fmha_ck_autogen_457eaffbff3c58183a656687010daa2c16cfc26e.hip
-fmha_bwd_convert_dq_d256_fp16_b64x64_group_o2_psd_deterministic.hip -> fmha_ck_autogen_d18727988e47264b42b4153dc82fc1a750f08db0.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2.hip -> fmha_ck_autogen_ab6cd5c9242f8278c8f3d9ce57b97d605c7e5a3e.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_deterministic.hip -> fmha_ck_autogen_0c93c65e5942a2f43f2e491547add02777dd2eee.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_pd.hip -> fmha_ck_autogen_d32c64ef01aa228277d031a74df51363f98aa2b0.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_e5c5079636a4a31a849ce8a5af89d50330a74628.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_ps.hip -> fmha_ck_autogen_ea62567e9ea16771d8445464c38f5a2931cb355a.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_c6e2da8b791d31f4ba05ef5f833fd6dea9e35f1c.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_psd.hip -> fmha_ck_autogen_f731289837f915e2aec1bd01eef1b3c1b099864d.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_233132e712eba8972ba444c604f89e01c5b84cc0.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_group_o2_ps.hip -> fmha_ck_autogen_afc4b47a6fa62a4ca5cff6a7e01c9f6b371d2215.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_group_o2_ps_deterministic.hip -> fmha_ck_autogen_bec30e7107c5dce3fe6aa87d83ed96da75478da0.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_group_o2_psd.hip -> fmha_ck_autogen_f4658c32d562f9d60c5ca1262a2e0df2375063bb.hip
-fmha_bwd_convert_dq_d32_bf16_b64x128_group_o2_psd_deterministic.hip -> fmha_ck_autogen_9545f95c1093c60f0fb6c794636f79aaeb53b733.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2.hip -> fmha_ck_autogen_e6b53fb8d81148ff384d31a703bb4c2e7a5a33af.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_deterministic.hip -> fmha_ck_autogen_7aa14aa94d625b33df1adfa30ef4d91769592608.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_pd.hip -> fmha_ck_autogen_b5db3d5b1d8af89381fc4b8073f84c5fa25fdef5.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_e8a9427f34bbf5ddb28a39161acc36806e68f2d0.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_ps.hip -> fmha_ck_autogen_724d1d4408196d611b2e0535bf8833652acbd6ef.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_a3ac4f93722dc314086f1b7d7b8adc687cd75f82.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_psd.hip -> fmha_ck_autogen_377b70f54cb2778b5ce3df936b477f775eea8b3c.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_5f20263fd84776f155519b3481be5e2c5b035585.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_group_o2_ps.hip -> fmha_ck_autogen_9745b04a8026a01828c5dd606d89d044d3ed1d99.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_group_o2_ps_deterministic.hip -> fmha_ck_autogen_a7784b03ad757d51c234fa86ea9891f055ecd5c1.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_group_o2_psd.hip -> fmha_ck_autogen_22105635385fbfb5d2f330df83ba6747bcb27f6d.hip
-fmha_bwd_convert_dq_d32_fp16_b64x128_group_o2_psd_deterministic.hip -> fmha_ck_autogen_3afbb5ac9048a962a60f48886728220ae6c2aeaf.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2.hip -> fmha_ck_autogen_429b82a27571ac91e3631cbdb7e0a58155abf962.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_deterministic.hip -> fmha_ck_autogen_dc818f3ce244743cb1dbff9aca399df90742a6d0.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_pd.hip -> fmha_ck_autogen_7f9403cb91d6aabebf081afae94a8ba397d8d24f.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_ca5681d4e5871aacef74bdba9e368445875252d3.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_ps.hip -> fmha_ck_autogen_1e7d7888480b83c78833214b32e10f37a6e20301.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_4018f690b6322588041bb467beabd8a7bc79a2e0.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_psd.hip -> fmha_ck_autogen_23047ea90076e3b0a3eb0586d49b9ee74ca6d279.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_5a216f777feec4752f5882677b18168225da4b53.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_group_o2_ps.hip -> fmha_ck_autogen_fd19d7614f2ed5da21a52ed172ef62cc07c9c01a.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_group_o2_ps_deterministic.hip -> fmha_ck_autogen_9893336a4b00b2a63f23ed7e13ec54c82d9e5063.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_group_o2_psd.hip -> fmha_ck_autogen_131c1fdc4206bb952b2fea675f24e3b09f605eef.hip
-fmha_bwd_convert_dq_d64_bf16_b64x128_group_o2_psd_deterministic.hip -> fmha_ck_autogen_cc4ac5a18f57f2ebb65f7e356e858ab0d59b2133.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2.hip -> fmha_ck_autogen_dde93ffe7fca311e136e42fbcd12b05c9fc7174c.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_deterministic.hip -> fmha_ck_autogen_7b67045d438a7e4b8f3a313a5df5a85f351c1be5.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_pd.hip -> fmha_ck_autogen_9689ecd7bf51bcffe9f5002959bdda41c50a3c8b.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_pd_deterministic.hip -> fmha_ck_autogen_c41b6eda4f250da059fe0c428428219ff5a250ef.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_ps.hip -> fmha_ck_autogen_c45a5e40f6a66bc5292a56e0097c69fe37cedfb3.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_ps_deterministic.hip -> fmha_ck_autogen_ffc6056d9fe125a4dbe08c1d86354e51f7daadd5.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_psd.hip -> fmha_ck_autogen_2995d39cd62f20622a31f11a292ed175abb5fdf9.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_batch_o2_psd_deterministic.hip -> fmha_ck_autogen_cb10303a0b79f2710eb7c66896d3c1f8b12c04dd.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_group_o2_ps.hip -> fmha_ck_autogen_81dd3ea61bb61de02667b14f5a94198f48c7307b.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_group_o2_ps_deterministic.hip -> fmha_ck_autogen_d3af8763f289dace1054bdcb4dfeda28b0aefcae.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_group_o2_psd.hip -> fmha_ck_autogen_e6e6b10e73733716e71ebf5a53703fb935fc5e02.hip
-fmha_bwd_convert_dq_d64_fp16_b64x128_group_o2_psd_deterministic.hip -> fmha_ck_autogen_e75c757c67aa23cb88e1aced6fcf36b7b28391db.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_2b3af90387f1d227119c5dcd4b71362940bbce52.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_e3015c5d50481547aa5754d042d9d7040cf1c7ff.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_a4700d87a19a173e84d64e43cffabbed52366e35.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_6af4c15a119e805e4407b184625f57966f8833d9.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8b17c082f249649eca733a8f0cdf9a1205c3e3d7.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_226662cf1c9900a4334d2cadcc5f5ac3ad355f05.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_d723b191785c97d284675f700a7baeb52a2eb791.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_afdab954fd111ec48721f25710d61c0c8affd8db.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d54ac01458df3f240e0656d82330f9de23ba9651.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_4ed6da5357b67cc28aee4afa9523adaf055c4e32.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_8c3bd4e029bba76ebfc79e6522dbc8ca0bba5dd2.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_dbde2ef18e2174ebe13a6e7c8c2a6b05a6612047.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_c363ee1b087f6b504a3dd3972b96e77db02b0582.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_a02a71fdd587e47ee68e0cc76c3c4494ce06c359.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_968fc75a7d102aca068e3ceb6111728c280fa837.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4a06b5b153ea6e8b1e20d9aad9d4633333fd98f5.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_bde24a8dbe6add6f2dd2beb48b1280f3a84a9b2a.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_415b183c50dd2663dabe3eb8b780913b778c54ab.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_258d747083272ea657604ac84867ecea17bd65da.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_2a97c457144cb63a9c6c3d6be613b47bd0df9928.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8d7549e66ef309e32779ddc2a1f14e79bae53754.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_5cd41b6f578f3c903eb9d58ebfab62eb296044e0.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_dc34b6ef496d4e0d8fbbe10731d4a7b1c136c036.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_4a9f3da698a6103caf25d785928dd9f814ac27b4.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_aa996b9c843200a2ec33ed4319b48106cd7c6384.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_1d02609fb803ea2697e2c2cef35e6f923d2578cf.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_58eb2edc7738d8d18ac359691da261ceaaf71788.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c35ea54eb6cd0f3756c462c66d9be956279b46ad.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_0f0c699d9c3b0ed62097e38ba05e40e815cf474e.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_64fe2db75cb20428856b02cd1cc8d7b393a6ad9c.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_38b94d76503e13c911781169fbc378517332c42e.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e0966fa1ff013e477b1706928de6cb7f8587c154.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_b9559dd36a0a4f5e068a722e285f485137bd5ef0.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_5a05b4e7782bd0e29ca9f6d33fc59d4304136d41.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_b9385db12001110c42eff6aabad935a69ad3afe2.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_c1f721a330b2d0fac13b22061616d7b10c0f91e9.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_37fe04467e87ec2110f60c7aea0cc9bf2ca07481.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_d4b99af9a573df50a27fccbec3fa8e350f1854eb.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_20588bcac681a5d69f252d7523a3681a0c6b6181.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_a3709e4fc53d2254a03ea7660b8c72d2f47cf1ad.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_47fe73f04cef91cd2a0682e905483968ff80eadb.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_ad9b99a194b59d3149842c15733394da275b12c0.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_062c8c3c1cf6c33af4574099e9b6ac54a55ad776.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ab1ca4ce061f7f69a250356f613cab00d1e2ac71.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_cd4efcdd12184211c74e7b3f2f30fecf1041ca32.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_1d0b822743e0205f60521d38d7c64f589fdf0f58.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_e10f47a44400de385ddbeb99475b717c5646fb41.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3e562e6c3af28b8478020ce3c3bf73c036001c93.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_1a99b2625adffa8215276bb88fc65bae944b846b.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_56cc4399c5567a9495f17d54c712cc9e65e57521.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_ba8b09f0aaa40a7c9ad5f0458b460d3e328f3c74.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_43e7c78e8f65be35e2753a0ad5123118555c56b2.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bed5a8c5cf683f6dfaefad72c2e2f5c2f2b2732f.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_73ec21ed6e040260c4f04ef68ef9307aa86985a7.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_3642b78913a853a62dbff8b99d9ae3fa458f461d.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_ac5e9aee85cd16903bf7b82a4ac10402b0b26e22.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5f954a393b7b5a7131c13d0c4578443f468a738d.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_78e945db4afa1330fe3978bc1bc9ae99828ae287.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_d4aff499ad527be5fe33b8e92547df57af26d40d.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2b8169ce4b4b9a17ac96fbb232e6a93f22071ab4.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_0a89417a043556970f72eebd48b4f3e7ac15377a.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_4824e1f8cda50f80988857611da766685da94494.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_dbae1670fac6812b2d2cbad973e4b475509ea504.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5daedab8931f2eefb649b91e80145cb71b63360c.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_dfcd68acfca68d1acac94f493e25be0ef20f209f.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_3511c54e6a6f9eec378d8b661121066536195d3a.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_deb9ec2cccab94920e40f62a1f0f094acd919d07.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_0fbb0bef3b388867e75d7a8a187b8b4b650a42ae.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ab0be5a2072b5e87f5ee58149688796b6513219f.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_91a6200e36944b1f11106c02f7fcee053f01ee71.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_1f81f8cce0d77dec9f977b9eeb0778b70a13fa75.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_bcd7ccdceb7baf3b986f2a0248827822a5f72e47.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_58762476c7f2bb05dce92ec22c0acbeb03676746.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_f4df1cbfbaf67705820f125b474469ad7ebab0c0.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_f42cf0e5fe479690883507028748b0cd3dc83cbb.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f682399cd6412fed6a1141296a7e4d42078f7b29.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_256ef175029a43e64164176d4eb212baf9d27bb9.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_3206cc121ce8955ed59ea3b12b858ee2e0cf82f8.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_d1840494c4fa78ff399c0399b3ad7ca3d22d4587.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_31c4b866692ba5c3d115482bef4790733863c1fc.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_b5c7fca1f76a31b0390e92d90d569fab94d4f783.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_dc3d625c5ad3e871f5a727ac946df642d988b9ab.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_ca4c6ad28aff1976c6dd36974ec3b339aa3090e9.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_eac353f963c52624cf79e82cc2b2c02eed94b677.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_29bffc159b0bb826ba489ae763dae141bfe8e802.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_9b327f0fa1155f2235d76be45cd22e3db5a69429.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_d0dd0165ee91c095a19ceddf08789e3576912590.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_5344427df3ae9392c4fc4c25c232196828e70648.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3f7315955f555768f24585a50d75e216c40f062d.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_dbcea8f7b5930abf76eecefce92d0db785d2df5d.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_165dfb45658df8f1ae8dc0738ac9614740f2576c.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8a58d4bca33c4c0e79141a56688049237d170d1b.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_fe9d98dbec5096a89b116f85675af772f023014a.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_d1c0dfd19a08d61586758091370acbdc6f267017.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_960ecb3013071fb65f2d5ed4c947c4bf303e5308.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_1552dc38d26f6badb7a9bcb5ce9124d54cc45ed3.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_3af86f458fb4dfcceb7db3357fbae0dc15142a15.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_74ba59d347ce8916a22b40e6f22a3c89e13db4d0.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_7344f96bed2f56793b1c2583485aa161cdf30379.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_ad989d2ce769f20e175fa88f4082c1c25fe03062.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_096e888c52d0f4a5847d7515fcc66208b1ff40d3.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_7cbe4562c51d6829ec5942e11035c452fe318b3a.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_621da34ee666903307d3a09b7a032f2a70054759.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_c64f4cdce32189065362a502105c31bd2d9d99a4.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_987f00dd759d9714693e7517dfaa8bb427294d42.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_1c2a2d78176e3f0a78e3ad78217e75a4430c0de5.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_ba145535e53899fe127987aa854f81234a9c51c4.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0968cebd81ade762c2f92fffc0153fa7a2b91eb5.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_b41735d250b5a16967281a5f07873b9cde3df4d6.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_fac5a0f98b94530befd634891e42c424bb86f0e1.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_ffb8adef0cef91a86f36872407fea35df90e8f2b.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_78e1edca5abe1bb3e7aa946eab6484b7bed806a3.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_88ed7f650c958a644c8031aeb88688b1e42458e5.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_ef2ebb4a86e7ed0001de9c5e607b66fe8877409f.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_f3ff73f82aee3184849d04c2364eaa45c6d0de9c.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_fb9477a613665cebcad781389ba7c5a36f51efe2.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_21f860d42fdc2cc6bd743d53ba546e332c22fedf.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_55ea83a47c6299fefa4220ed88f7a8e1dd938215.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_80987e2d765efc320eaee813607c94c80ee35aa4.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_288458c5a0720ef152848713119ebce6d76db6d6.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d6149eea92f2c40c11de3b778102fcf9b6a006b8.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_7b5680f97836be4a369802e8115617a83875703e.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_4347e039c003489dd528faf5d710e687321a3fd7.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e5b2bb9f8466de1ad5210e4c39ee7b8ecacdffa9.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_bc6ce17223d8d83a64b8c96ac88223e4441a4692.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_fc1790325b59bd44b0a5f6cf9723a25fd845cba7.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_662767e588220d0dc6137b00cc1d8dcc91e97134.hip
-fmha_bwd_d128_bf16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a3dc780b17152f696f9b957432c2eae8fb16e85e.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_9a8e04fe9432a60f86ff0369e8c1851821074a04.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_835a906031a258c6362313eec783678bd8125c91.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_ee8e709eec7aef1fa681053c6d2969a5ff18c45c.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_8d079c1eb36db8461fa8b861c56760afcd97cc34.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_64b3488ddf3bb1a4870371882f0a5d267bdfdf73.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_ca3975efd767ddf7c12e308d948bdcaf0968493a.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_82ad0c0580516485ea432d98f53e73f6dfec548c.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_4306c6c37cf472ad262f53941611b5e60072bdf6.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4904c5910a2d0595b39a3f87652a9d1ef4fcbe80.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_f57f84892e2a8496169b7406e63b0d4f5aa63aaf.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_f24f26e45d5cf567d29fbe375fbf8abdec39186f.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a85d35b2fd98742427930eb536e346ffb005edd8.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_19df4e13108e043361e9528b71df56f04f696a0c.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_dbb06b43d5d65429e23cc717448cf1fffb0cfd74.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_0ef9b9413697d6f4573c6605bff6f58d027c5016.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0b2efefea81036641561bed80c75d77651176f74.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_27c2000d32c230a57a6712f27bc0fba02722f5fd.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_ab1d7f93427095e39bfc1d986b3d7fe54073ec75.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_7dfe21ee27f8a0ca0407ef0dea73cd73ae6940db.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_8007bf7ae1b71bf8ac4a793aa519ad333aa7a7ba.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3937d9dfb68351de2942e32f35e2ca1ce71edfa8.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_de1ff66d2aeb47d2fdccaa4bb6b9d066b380c99e.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_5403eec1cdd216d5c4a7ba977e2ef92a0d7fcc8b.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_358399e756ed5026baf3ab78af17489dc07b9532.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bd064e302ff5b983dbdb4ccf51383fb29ddff44f.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_c11d68fe766fc753c657362673704005b538660b.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_fbea85b766bf0c918ee0baf24dffc6a5563d5105.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_92f9ad0fb65638cfffb3e7786f2cbf01d9585b23.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_0a55ed15ef58c941e06dda890aeb530e28eb7bba.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_df4bb75ca79f805a81fbad750ad22f6d22b0d8ff.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_7ab03a62e064864e1e9c1cd506c1b2e1786a777c.hip
-fmha_bwd_d128_bf16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a189292c81a18d21a2921ce6740f81ebf4c046ad.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_c9312d7159369d13f3148a6f0882dfad6921ceec.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_0cdef49859c80c6b3ba18eb2fb4c35c72abc1cf2.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_ae87b1d5c50606430b544ed650d87df24366e7d5.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_0a92671b6ea99891c0d69b1c793f4d131b9a82ed.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f4a6438394dd3427f29aa0bbe58ad1f797c3c38d.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_fa85f869a92f0482605e52019828244b12e12b44.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_c2541b6b5cf27de3f45f60671d36602f07ce1783.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_0595316f0dfffda03e5296b959a49ec3f3c48d67.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fff7aa57cca501f221077124359a589b3a6f9d0a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_358d28c958c0a831a615a4811d13279b18db09c4.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_96f1bb85dff8c97846f6b2e8796a6289bcd0d9d3.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_14d4630876785655bd4950566e81ae0b645c0d3c.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_a48843d844f78690c7a45b730652f0f763c595c7.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_3e143d88eaa0d9cfea856b2f3a57d1275a656627.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_18ed7195a9443c84956c3f32839cb3ab9056bdfc.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f7035f4bfd8f2f427720a07e3c311bccc1dba683.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_f87790f260630f312b84888dcbdf849ce130ae59.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_fe97b7adcd67ed9bda8831d1f3f1ca7590c6d251.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_b41a30092e8138877c1f6c25656e0f8ae2c2444e.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_af06c0dae15684f83e15722a4c07342af9ea011c.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_158d5ce564c3ae1eefb54e3d41dde2604560ef4a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_49f5017cc0f5c8c8dc71492e7765cf729c1f225c.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_280bfced8745fbd9266207463fb41476dc23afff.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_eca613eaa8471ad7da66d2f8f2b8e07f6e02b467.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8e1b48a28b71c7f4c78eb14321b39951a7c5e903.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_a1d6ad9de7ac7993ae1923a2ef070b7dacb8c563.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_04641230fe9a50a221047f7a1df8a370f72805b9.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bc1ae1dddb8cc5d78196da6b26ebe66c1ce7e567.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_e8d9b65558398c0c10127b560807578ef117d7ed.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_87e3a06266deda093bdf28af82d8666066157fc6.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_0a672fca51de618e3441cf8764e8e83eb782f2c7.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_92d841e6d783bb46d841aafd9027f92dd1b61b88.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_01f74764c3c3284fdd1b67d0ea781c2261ed0de6.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_feb5e77111fe1e20bafdb83a925b5faeeb6214af.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_26d77b228420a3ead919474ec9c6fb2800f86890.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_4fd34faa8b168e2ac7862641229e6146d3e28aee.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5724d91c1fd6290a6cf8d52a3801ac6b921dc7d4.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_dd11806cd2d3ef1127f676b2d98bf8fff2a1e5ab.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_aceb0641213e9a45ba48bcf72bb23845720d8b79.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_c0338fbc05f86270ded7df2bd3e2758a03961b62.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2e8b4260626beeac76c26dbcee3cba1457b30e99.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_4e0a88ccef04e81b8c684b695f7cb4310e448915.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_6f31b3345893eec8ed1ddf1d8de2512b46ff6187.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_83d920a76114c63156740ba5dd6f3846c4b21c28.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_86fa51b8c7a2f3fac5cf4cd2951ed2ede5c35450.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_e7b2eb64b66d46359fab44333c2c484f4c9dd5de.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_81acf1d17650712b71a499bb66909bfcfcb6aecb.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f1ecc90ad7b86791a9e6f73a582aeff30f393804.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_f01468c62c878295443981662e037ec5213cf7a3.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_e2deafd2f36cee29109fb824e0135407453adcfe.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_b1766695dbb790bd614b83dc7569ad449404cc89.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_784c35fee4d372123631312f1051c43e1fa12378.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_38bb367362fe2c4849ded728ec5dd00969ce188f.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_9afe4b6f3b901ff4af81bd4f1cd8ff19f09d0b07.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_9ca3b1d36d777213eb381b47871bf15dd163c994.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_d5edfe3e3dc3008b928c8e6dbd50784b905f189e.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_10c24f1f9009e46afa3a59193784cc2575f79056.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_7dac5d4cf103d658e129673549549f1276f134e0.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_c8dbfaffc8a9b573f194f9c63f1175d9725f8950.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_566e26d4969bc6bbe9b092bedab11cddb3360c0f.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_0ef309b923172f4c0fb38d9b9f5325b33b4877c2.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_3bb3b682eab96e4e173affad75b9d8e73f1dd690.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_f92e9a82c879051d6fe3c42108f8a574187704af.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4f44435491aa68acb3217b0e693232c67641a2db.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_6082d55544b5280b49b071ea277fb1827193fa2a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_81bb8f13b6f20a72c9ce6d0b53f81eddbf05f1c6.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_1e42736d4f677a59a172bd6f162616a437696351.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_b9ed0a64deb55616646ea98b21a891c971cd98ad.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fb2fbb135d59028afcf867c2cf08edc323565528.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_6360621af3f7e1e81a8be48fea8d2750fdecbbf4.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_097b3e1dae9bfb2e89398706508f8e01966fd4ea.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_4409f2a7deb027e864afdfc9975d3ab93c5dcc9a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6d307974bdeeef95cca0d130ebb7aeb77fb1b6eb.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_01ee0083f6df962c4a754cd3295b1a436c590a0e.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_c0a3c4ac0a50bb9b7ad764929dbee98c856b1210.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c27b3026f1dc3056dee3a3e64bf31c45683607c9.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_5af96b404feac271dac8f4190180754480d3ba80.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_f69878f4ca8cfe6b8d8748766f66a1ef8eab20ad.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_8689126a7eb09d81baaf8f99dbff8932fbeab3cb.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f6856ca950bcf173571766c3f04de4163be0402e.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_d036096f49a89730f8af7e75457c88cb8ae64165.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_03ff035717140f7385282419598cb4fb2881ce8e.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_de85901d66dc04b1143bb6404445baf65693b781.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_5c742b9ac6749f189d597ac97d46d35189472c50.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bd9c47f3305e47db6ab6bc627fb3d80269633074.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_d82773721479613ad72e334510a248f1436b38d6.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_cfda56a4eb08b803332f25bda6209932d9624acc.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_328a311bafd1c153525393b252e4170f8aafb370.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e5935fbda313d3518f142f43d46f56c600f69286.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_48e9e858abf6f77489f3fadc4ee81edacd26705a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_f71f96ce4dcc7f789a8ace73c230c203b05ff6dc.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_01d12033d59ce2799a2a024e5d9232325ccf1320.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_5854f09511778dd1779a839b0b194896070f69ad.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_7237ce5f3cf13ace3efc0b0227ae5a8c1fdfce1d.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_3b4ecb47f9ebe8c2784976c3e9bbe4834b475cf1.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f18c74becc24a93427d9c0838784e9b6caad6e81.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_c4c6c405cefe204824e8fad1b3dd34bba87e796a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_41db3f29d1940e59dadc357c040ea37a6ff208d9.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_df4c9eb48da49a61957537270d94e56cb4e426be.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_6018ab272d7306689c7dc5a6d5326efea1471235.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a421c2ed6b295c458071f1988b9d6f7b46e8992c.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_61a44ac409e914c12281f1d26e5b52d8bfd0df75.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_7e332a6aeecfb12dcf70c69157fd3137343fb9f6.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_2e43e401abbfb1b6737e4dc822f68421abbc648a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4afd02981f92fbef6277c1985cc479c12bae9239.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_8513d96a66a4d9fb8dfc84afba7e1d8c200248a6.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_c4dec99707511cebd9188d216ee0a148d729b470.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b75843bb13058ffe29251e053800c509c7590544.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_6eca9cd905ea8b0454cf9564643894682b08cb97.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_c4b34d3cb673447773f6da23e9cf52b98e99f718.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_fbeec221cd63adaedceec39db41ea942f99f5133.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2d7b637e0313cb423b22cd8844cc2997b3ff73e4.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_8fb224b40a7be7db0a9c5c08cc5ab05b526c14e8.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_e28fd64c2f2b27577109a984e6ab82f5f0fcb296.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_2eba937ff6d0302ab013db7349d4feb914107f1f.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_4e79dce18e49ffe024fe4cd0693ad3399f5edaee.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8cdcdeb845e7bcdb89ef70ab2a97157d4db3cb52.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_30024440e780fdf9ec94deccc85216d8bbb5788a.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_c1f40c3421b9ad8cf43940530ec50bcf620058f2.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_7e89f79217037e361bb0909d06534e40f5026b4f.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_44564dddf8b492d80be54854abb8d1d831e42679.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_7831ce329f2a0812ebb1dd103ea4ba8cb7ba531d.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_b7a03ab0b7887cc7ed0cb40e56360a8d36c0bb8e.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e986d5f8d5591f3e0f1cdfad19c38c420fd93023.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_076b3beb57b30afb30636f948e3989b346b38d20.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_2177d95cdf45f6fec95d1812f2ef183a75259e38.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_ff6862dbdbb20bc63a650e1f93e9ac169bb702b2.hip
-fmha_bwd_d128_fp16_batch_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_649336d59a8b35919e593217b6fd4314a04ea359.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_14d11aad7b666f500f68b264a2fcca6dfc5f1a05.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_4d5f3cf0f78f73df79665c26b20b0805615e1b04.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_4bc48576f285325345fa1205e5e7e01787b74f71.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_e7c0a99e949baa5f3a7ee2d6e84427982f82f76d.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4a2e6b05e7e4de2cb23d815f8b2c8adf22131c0c.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_0842c4e3aabdf55405b3ce09ce1899245ddf11ad.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_9ad1f99284aafc8d7908d062f179a056eb314925.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_b80d0828ba6d24ea3c1a97bd9835ee937b4b32fb.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_847feaf237911478173377a501ee19ee325b012b.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_1a8da3e6ab050262b659c801ccf9a14787d7f176.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_0225857454eaab2eb664aef7a0849ce12c32fdf9.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_80a72d70d80b66c19e85daa00497308381050048.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_8b9043572cabb65435627a3faf23b18d039bbcd8.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_378759ae25465c32960487375828e23c5f1ac869.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_83ddca2c6ecbba4314c434e7471ffb8fa642f936.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_40db688a9189e1c47c300d474df946a248a63303.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_ad091c69d19b27f7ad50ef6311532ad8b642a9c6.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_5e735b12d130ebf849ac5d6752e413ecf3e69fbf.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_2c77bd7e89ed832cc31b2995566a49bec6e4cb52.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_133c51948cf8584900807998da14d788039f53b9.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c29110dd501853e87ebc122dd1971b0bb1bcd92f.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_9a9edbe35a8fac7796f00bde836bd547044770ea.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_ccac6c0e61b65c9422c7f30fbd979031698370a9.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_0d13a4c8d169877da6408584dc1f20a6f7c5e3aa.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_744ec604c577a27e0aae5b39711a9e2eb82801b6.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_43f2156a04b18bab55af60e9357f28d8a4604e8e.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_dc9e54273c0ea2358fb573a7d918aa7b09fe07f9.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4f0aded9d1baec3125ce8e176248cb146ca580fa.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_c80dce1a17d073259250ec0c87ade69e639ffa8e.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_e307a1b0d5a8f94e0a0f4032f401d20b4b643523.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_05538339c21c92c53d237865d72debaaf2ee5075.hip
-fmha_bwd_d128_fp16_group_b16x128x128x16x128x16x32x128x128_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ec3deb1382003ac010d9bc1c59d1878d3ec7a727.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_1f7faa0b33a9aada86f032174afd40d18efa7715.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_4462b192a64efb60d5484798526278ac7a0fb9fa.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_3a2643099365d0903c799585f41dc1a525ac9f9e.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_555ba79201a585bc091ccfc326fd24e851d1eecc.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_153e897098539c3466da9d7a37234daf16476277.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_38a5ff72f22e0ad040a281e66b1aca0bf3a2aadb.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_4b2e7f96b095ebfb66ecc7a75752fba2a63e4f37.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_0fd4068ea93fcf4df463e3bf3a6898d23b65da7f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2b823c3b99e7c8d1cdc39a5dbc7365a383bf9ccb.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_3824e97d5ecba46e06d5ec1a9456c810d80227a3.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_a5d4eb673bafd81e3a0ee213da4603d88b8460ec.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_40aa64439b80ff8dd12498b3e5f6b625da16e285.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_f3bf7ef503bb026258b3ec3d82d3ef1443046964.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_556cd05288e1666f5c67fb87ad02ce660e4c589c.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_fc030b61ae20c4b7d9b2d10930a17e01e9e93328.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f069b38b26c30bc770f74c856e47eb498f5818e7.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_fc7b0916744b593435d8e1e7b6d874d760cd5e3b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_abf92a5314fd33491b5eb6ebd2418b7e0d5db774.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_d41b6a64dd181f2efa65aaed03a3d229b3566c1d.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_80bfb0e6032892cc58cef4dd403f305a5b76851b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8301bfc0394936a68fa0098580f06e77c88ebed9.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_e9b53fa68641f45baabf40b7cfb8b35a9a1b9c7f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_c9fb8343e623e46f01893a2b61345d1ca5928671.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_320a6196b662a1d3dc7441a9536d825dc356b95d.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a3d7aa46528ee74e2bef1e87c1feceacfa55e173.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_c59937be2b9a13d6520fdcc922e4e75c9fa085ab.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_e477abef05ff37ec27705eda51896e2aa3a04966.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3da8c31f6d5bcaacfa4a21aed4d1d3caecb48922.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_6d40d762ed576832b3a752453e9881b5fe6d2650.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_3c1454ffc1418dac641f63671e947d9f550b1f0c.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_96c129dd4c798343d6f78ab78056f0faf2f1c9d3.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_242013527a0266ad479715ee3e6ae01c45de29d0.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_2dfac5a83def98340c8786d55a30a98ad68b9eed.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_ae51b30c7e1cd30e550187458350c8db7c59a9ef.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_5e0abf4e2b6be3e2c555c2134705b9dcaee617ce.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_7309c38fc8a2d5ad6efd449107dc54a7509624fe.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_245d90000b55ab8b6055b1934880fc6c4870b34b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_0b9585ba1c10acf67115c5899b3546608541820d.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_8e431313fe082958d31b68d2fd0d61df0fe56736.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_1db03461737f1e359f389a8d297476f9b60faabd.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b4b037a2e262d11d3ed7d9feeb41b9e05427a739.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_c919b8ed877d4244d01a17ecb948b459e361ff24.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_64cf03c0aa3f1b2a7b76b4e3418eb5063b982a29.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_1386cd75411e61a8dbbaf2b916e62f4f5f99104f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_6e8cda718e10824956f0ee39bbb0891eafa45a7b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_2ea394a09c8691a534ad2219bedf73724b6dd5ce.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_748a3d76e8ab73af9a5d2302d33e3b1d1b866dd1.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e907e8d1089557dfcc95a05160be5092e9119a53.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_c4c3425fe683d35dc3335db77d183ad1620b7a92.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_d04dc4ed02eb42c3fe303342801ed3073a0dcb8e.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_3ccf0a9d5a5451da5dbf6075ccea45e4a140550a.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_d924ee32b178b6bffa7a71603d6e2818f66177a5.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_aebd5fed34ebceb879ae3dffaf58c7c04ab5fe80.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_5939e6610e41aff8d1ccdb66d9e84d3e48e8d379.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_b4bd2d206ceb237ed2c51f58abb5cbf96e39d07b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_e56757fb17f5e94a6ba1fb14540a68c36d571159.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3076a6de0e2612279e0ed64612f7393856bcc9ac.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_ea6a6d4cc262ea838dbb83ee747112f95fa297bc.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_1a6bc2762b95d550485aa720edaf71138d94cd07.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_614a9f10ebc51bde3f580ef527c17f89489c12c7.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_0271bd8b7c270e1593871b638288a4923342c446.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_4b74439f42140cdda9bb0f78d995d741212a35f4.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_d733f4c03e338ea7c6d8f759c1132499bdcea059.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4432c5214c4d40c54ca2d02f0d4785c6d6902370.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_1f13a6d0f8c798c0c4ba4ad202d081899fe081ab.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_a1c71e7d33f0597fe090a3524e33e18b2e562680.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_e13b86fe4e153e0bfa8d1e75f3641fe32b0c5149.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_adae2d4f8b2dac799e03ea6f279e6ecdf66f5381.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_70586668a61ab88bc46b763df8f1c2ea52001ea0.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_82f0f3d71108dcc49234a258f0f3b21ea2123cc0.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_1de2f97d49f015b9af0b186801e939c6f357a0c4.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_9dc424f0e192155e3c4e786e5b87d5a1a3e6c4ad.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bc744db85d4237ee9640f1658e0caab7648e3bb6.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_e8d8fe5f4f8641998b8b805a20b2ca92d019ee59.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_549b6956eaf678f7eb901567d1a515eddbedae5f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ddcb1cfea1b0dbe50a02252cba99428fd977527e.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_86d73393d0d8b769f30222f7817563a955c36dfc.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_249668a3212cd00edaae871758be30a5a1fea589.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_643b3798f11997d33ccb58d90ed6c10d5411b735.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_adda7ad787524e3e47dcc1b65c41b2faea38f55f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_5d7ed4c885fb32a0b548186e56d64bab98071d30.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_77a814291d8f01870274149b9d82fb75921d6e20.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_f395bec57c3b2e6e169134dd8d20b287d7405134.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_0f588dcb2ef86677ebf84e406eb802e9921d1f1e.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0caeedaa7d50f1741d618fb6c573529eebb075b1.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_1e33ce1fa113b221e5303b4093c2c4e748ce8298.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_ee974931e65d6b16b7c868d462b95dcae20b7513.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_85960fe542635079de5eca3c7785890cd4740005.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_089de13222caec1483207d4a54249f8da4f9c151.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_ffb5b7349a671b182d73c8016590f26fe06a4cba.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_768c80fd3ea17813df1bf19a158186834fd00780.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_597a0276ec419f18f060a5186e6bb703ae434ac8.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_fc86c13e933cba40553ffba31d53aad27415ce4b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_31c3760f5978baf9780ce4587ae4c768af0e49d1.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_92b0770fe64e3c60b9e56170aa88bbf74802a813.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c487a1a9933239270f44b1e08e1cf5323521c089.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_3a1dca5feb864e8981387c2d07e62acef1730aa8.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_96caa2056d99eb67ada498e287b4fae984397691.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_683e8a33fdb7053760c9c135002b0a94facbe015.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_7726be8909f631c04d4395fa4ffd03a736f447f1.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c197d1f050f42d82e6851fa286db6f81ba197f40.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_d3a23ded424200d0c6f06b1dbd0a7b7b0e7b5d9b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_6ff4605d82507fc4bd6e96095eaee5173ea41973.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_71e3980331dc4bcec6ab6f4c345c7b5f71356979.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7e9c7feb747241c9c7de2adf3a19933a1c4c0995.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_1a236be9da05a07d11cd28034d90cdf89941a172.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_ab0c3fe9529e24327686070731d0ac3ada76245e.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_66be70b088b20fc8de464167c35745461ddab640.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_fb4c15452f9155c5966990f09432e5eb7e28e785.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_17b9b96edda151072215502cc2b606bf1f6f0b03.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_f36aaa63ed42a578b953ebd614318d44cf44e8a3.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e578ec9e09d3b78dca6b5bf0be1538657f02f319.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_09513bff5c1da6aadf11d2e8272a422eabff21bc.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_f020134822739be6fa0bb3d98e9dec79f025324a.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_7a13d62a715fd717f0d4101f787349cb49cbe70f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_d40569ae9dbd693c0ab3d6ba69704d31e451011b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2c808da5c2514806c2953bb77d5692e5d7c97aa3.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_bc79e255d25744725e2a9db9f90d5cc2b8a0e0c1.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_84dc4af43de08130a04bfa06df9799b6e9e96900.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_006c417a52a1bd7c55e45d111483d26f4480caeb.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e02a198f23c409b715761b702d7b0e6e5992701f.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_468a5f057fd5cef2df5f919f5102f47e86901e3b.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_16047b5544acef40e39932672cac6f562e200948.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_741401abfbbbdf0dd1d62df8bc3e85371ead71d6.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_9009b7d39346537aa6c4a4e46b81139f603edb60.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_75c38912947881caa14b3fc7ab7bca317e296dc3.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_1e943fcc2e64c618fc1415b3f1a0db4d70aa8494.hip
-fmha_bwd_d256_bf16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6d470f5c6fb81032fcd7974180297d4bb2a8427d.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_aa1041530f794c7b8dc4a8321ea0fcdd338fff35.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_ec9f63a538940e5ace02ae5b5ddc01f730adac4d.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_459c8fb6028991321b09a990c2188d854d940268.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_a2a715b7e9c1a576f011dfe5769c5b392e984f82.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_64c3c1e3dac623f07c2dc1b934ccb868cafcb38c.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_ccd0b777df1328bf24e070ed4cdf8615bb2199fe.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_6dd707cf48a17d31abef94215c5720419faa0a39.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_687f4aaafd1a5b9ee85aadc6fab79ad0c27a2ea2.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_aebff7e6605b273bad844b8f70ef031625bff48e.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_cc127a63d56099e08125b16939dac82f0173122b.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_7838849e57ee9cd292e588f587a8079b57becfc8.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e638053e01268a4c5883620fc6a9901951e2e01a.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_405e7efa263223148318ae96bd1929b382e994e1.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_4c69d06e3f32e3b6d28d3e54ad764b472741c193.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_64a0ca185449a49fa485892fde6af745ba758167.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_af6ccfa11add1ae49888337e84d9c446d2f67da4.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_a487f617c4b84c6a0328fedac750d41dc3dafe27.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_6e6a4475ea795935f4cbf2dc0ac156a33d754587.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_d95835bc6f000d3a3379bbc38d90e83dcaf867ee.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_2c2e75e6f659a500dd3cf2cfd65118f111342119.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_28f2e2b108a53308a0cb6c123c8d318cbc2eadb4.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_a65c43b870705c780d734f9ef063f55cf8b3b52d.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_8fc08b4f3959a2375ac03f40c4ce12d70cdc2d80.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_a673f35edd69241c6b921d6712dfd064d78ecbad.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ab877ae2a1aab04498bf2b26b3fe99d6488ef151.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_7601e6aea44b96e94fb019501be6b102c6e6a654.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_4ef35d82ceb4af2e07719c16109c6d72eaedce67.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3c64c33870ebc329921cfa3867d58b1857421f65.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_526c89b7a04758b4badbf9695b316f877b8bb053.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_b3da22d3482738a8474ae15e8e5fca9020c4e195.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_f672bf80a78885428b2c02e522426470653a7351.hip
-fmha_bwd_d256_bf16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7c19fc90e5a9c422dbf529d2def286f47dea0f50.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_76704ca28a4877a1e84022e022614709adabb280.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_0029076f83a3dc695a167beda6fe19230a2b114b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_da29a515d14dac02066bcd4701285b9916b43cf5.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_33e7c1e5f41a451c7baff54f7238b220f1bdf8a1.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3c38bb80e9880335faaea81985ed5d0e713ecb08.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_77d0223697ed41c4c2fd8830f8df6e5620db547f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_987a617fae00fa90a1ba60937b0312c81087c19e.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_1a6785392af35e27d6697b584cb6f17a766d3fee.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f3fd08d56f8a9be1a8dd104cdb1ac58e283b5064.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_73d4901b8ef034590314048de7223a572d61ee0f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_0502e718337eab7d47aa65cea7d3c5f641484520.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_618031345ea71cc17e458eb97a559b7c94d3ae43.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_14c4ebd1792c781d219bd21b691b575f64635730.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_56de9a7dfb1201b56528740e9d8a07b62710fcaf.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_cd0453a5c3828c1358360f31f5d3b7258e17fdb9.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4cb1861e31df98bdfd731efc3d335055090d83af.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_be8ec1163a01b9cd9a802d8b44669e8770c20234.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_f0cad48d9bc80d58705ea60eb2dda4baad68cedb.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_ef7cc2aa1ffd38298b52764a93cd1271b4d92f8d.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_3408103188e27b3bc55dce0c1716c0b4d32d6494.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_1bf767e7104cfc8322f26df35907fbf04b8948f3.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_9594816877815bc0294610ca24f986fdccdc7c6f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_d9061c204d8a85c974676f4438994a0be9d69a60.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_becc2a4d7ac045365300bf8bd45fc6d3e1e1c8b1.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ddf5339054f47d9ed6cc7f9e66ab21ce3bccf3db.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_b01dc872c24db4db0c9179fc07e17f41060390de.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_84e8ae99e184013739019c93d07caddce532382b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6a66604bb15f97a56847a7c968dbe32d247cbc13.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_90e5c56e92712d00092ba102a5eb5176a3e5d471.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_8352031044ef2e4a22e27ad04ab5d2c02121faee.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_7dd260849b86c46b685955cab54ba07d49b47954.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_afda8f46b5ded4c2aa9d722fec17b75004b59f7d.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_98e484adeddf3394d8d7693b808d83b64c71ee69.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_cbd571f4fe576fdb17d5f75a558cb6747087c7f2.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_378bf438642e5d863e31145ada2a0688059aa5d9.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_95530399ad7b43d8ce2c89da24c71056f2146b18.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b00e062055933388e37525df5766f3c14cd3538a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_236b3eef02b904304348b9d35f715b639d63218f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_069c663be0267c009be4814e9e4e7c13ec999411.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_a017be7b8bcf303b30a147f41346898acc5fab7d.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d20d45aa85c0daa299da98c277cee826fe67bd27.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_b34c1ce348c3d9cdf6bbec9758de9d5fe94c43fc.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_1c1b0f85e085dd0769c566fb16aafe5ab5952714.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b513834918d5ea789e2db21abece7c2d3532a7e7.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_0513b2f3bd8ad51315aadb7f63737201898adca8.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_4bd4d46397a3749646b232b306688e52b8c6e584.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_f12f1f1b679cabab04218037ef370d2c7e1fe332.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d623b36cc3f56d1001b2d3abadd8a5628fefd014.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_3f5e01b4f2ca8ea10898c39d6570bd74e85f46ed.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_a5bdc110955c05c6c6ea236a6f60266a4a6dce5e.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_70c8e45f6ea7cf5dba9eeadd0b19481d9f5defb7.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_b5371415448fffffd58bf014dac9f4876153657b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ae4e80cb185759dd9b3eb3c67c239964b3694caa.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_096863cd93d1b105a617d0daa1d4f37d7fb6b893.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_ae8d0bdde763e617beafc0365ec4a3cd11df6c55.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_f7cf08242b3fb1c643d4149bec985b667b9d28fa.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_44c181996532676f2140fd026707135144e9d37b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_8f6e463eedd3e65b9c79feed3cd92ad8cbc9f036.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_9638c9618dbf2af119e37596f7eb0fd3f8d72748.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7f80d44e82e601dc48d4c8b4e710ef7265894b6c.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_85908fe6dc9c629c82d6953081b10021e64583b1.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_fecd7501265b4c4dcf015485e63e2324304f70d3.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_3b508b92f7e123b21658f6e17d624ffa87831fee.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_01e2428c5447aa9a78f79f73f31cf685c586872d.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_e088f0f7363804cf5403adef70828ab32d09a02a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_f4900c0a5c0d03dc17d7a907ab40652d9920e756.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_cb20538073888bdb3174a8e9c32d7449072aa753.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_6a3f42d5c9ccdd3807e488b00f02bc6ab5d8d99a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c9f1e7e478a2208c4d32e2d7e6abebdc16bcc5fe.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_8457ea5726149efb8778e6d90798b8e48288fc9a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_37ad61bf8427a26775969f8a9166fd0bfb7446b4.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_72abb25dba0c48b380b2dabeb6ab7efaa706d180.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_1a5e18f6333ed2cce509f07cb8bd5868951d66a0.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_091cb49c1958fb4342d79f367ea93cf2b472f785.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_a93324ccf11b273ed20fd960c61df897c8890b1d.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_906fa8bf5e992ddc25815486ae9c24d8bfba7227.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_6ef5803b33d97db72eb8a8528aeb3fc956a938cc.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_a0874fc5ac87a1ec487c7722bf3b1bdaa924ee09.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_e7ae1294b6dea5c8b93c2b814fa7460c4047105b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_238e4c1ca112afec494fbe47a85b553302c43395.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_ab09941bddfa9d61985b55f9b6bf0edec9bb89f6.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_48280c91d7cd8712fd533e246a6b0f758834abc9.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_6a95543aeed81adfb6d847f78212585a36122ae3.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_6767cce35ab784aa42ebcb75af7305bc38a8721a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_31b807c48c472e9b1311a6037cd98e21d6706889.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_dc4d27535b9570b8f4b790470a83c1d0a9a2b6ce.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_ab56e886d53a1d88fada0f10f00b9f398dc54568.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_8adbdcd28cb2f078f89adf9aad2b3d4a0a477823.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bcf8836c8cf932cc2748e313885003f0e11a887f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_2af6c5be53732eb1939a2f93232af7dc011dec1a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_da9f6e1d59132fe96709490af25bd794f267851c.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b31f56244076c501cb09b4b90975132cae4c4386.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_f9c58761c927b222112cb5cb6c9acb5d3c915785.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_041a0718891596ddac1fb0088637029233ccbe60.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_9801b25e0f132d647934deb395b62a3f70cc7c88.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6376eb68c550b50b9aea42a7a2cc3bda186b0e40.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_810dd4e870ceda3ba9b5f0084a4b025b2e609d57.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_a821661d8280c6e9d27f2c9ce1b3c855387b5a76.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_4be4a98f150f3f9ab6f03b5fd0968c5454565c9a.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_96dee49ec6755006d67f0c30c65f50558bba69b0.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_83d580a612af85533c87aecdd7b0345c71b75980.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_451fbbdc2dcf2ec81efce34673ee6c425cc16ca2.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_c4376ac8d82db1bc25fa273a80dfbf8b71ee5e2b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_910cb8bd09d287a1566265eb1e8894fe68d3cc81.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5b7a4ea3bb8905a22ae97a94c354b1cbe38093bb.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_da07d8b5666423da30a95e3b2cabd3839d200981.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_5bead6be6e39ece0e5d44335083336f7f546d2f8.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bcb6f0730fd09b4c6c60913425927dfdb8f83d82.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_ffd868d49abdb769ab82c21508d655daf54b8a99.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_d9c3e27b522320dcca5ee84fa534b03aae2bfea9.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_c323a4d1f24d59bddd20ed2f2fb6446627b0ae8b.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fa16fa84278b489af253b52839786f94aeeac36f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_bec9e4c0317e8d351f60258ed6611fbf365c4024.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_13d5f2ec83b3331654e37ea0b44d88cd98abaa37.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_fd614df484b263deae3b3c20adb0ce7b62eaa651.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_05e60b3ab7477f9edc8576a8bf43e3a62b8d5ef8.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fccabea88b8e290688c1b360875d228e6fdf1624.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_486f6c7c7655c34b7b9973ff357b0813f0a3fd7c.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_3cd7a9ca49c1149d46f6b05b0fefc41ecaeb6ea1.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_5e62968de58d9df7d687d671f37d63393f189321.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_807545400aa6e70ff49a5f38ed6a218a180bd87f.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_f5803aadd93e33567aa6b23100ce4fbb6c040dd6.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_4466b6c6b2ec3acb40ac1cda432efa1e4e62d9d9.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bbfd025488e52b97c04995c4c5faff371b77e4d6.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_b298e213f927b518c693660110f08bdd94990ef0.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_d090b771a4f9750132f549c82a88b4ab00dce5c7.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_9068ba8df8b0e977e9769f6acf6cfee6b00b9922.hip
-fmha_bwd_d256_fp16_batch_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6d17b92fab5bee7717bf9aff6a6bef7cee3816e7.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_dd10bbf37503bbc92af82bc3487989b41b20ca85.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_f0209426a8e6bfeef7d8ae7b16db791888142298.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_e89bcea4393593313d18a4aa6dcb44cd75bc828d.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_e34b7e452a4db74189334697e3a240ad68085f0e.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_615430cb65d8d540836c7f12b3367abd3c8e63d2.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_afadc4f76e237514db0bc0203102297b79730bd0.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_249e6b93baae25dff97a0bc9145a8d328ed3f317.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_c806d7803d06ef8aac1d5caac9f36aafd47653d5.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3163272d25bc2db2ffaa1fea87648b45ee68d408.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_b9baf70220079e6d4e87eb01a7259923d8a01e29.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_c5fcdea177734366d3bf283317a65cc3fffda611.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d25ce4b3e9cc392ceafebc7fe3bcbe05aaad4bbc.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_3bb129e6dee6848043dd0e8fa812ae80fec4d014.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_7d2f87c021e0b6a27b2d7e30351fd50f06414b5f.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_f4c803838f5644ccc6f04f7c8a6233fed0b6639e.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2c82e3c4e445e1e02f14435e4ca01a90850139a4.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_a21f3637624762547af1292e1b85e640b1d329dc.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_c9ba0a3369d4e4eaea1c902a90e6501f232dd57c.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_1914250fce818584291c69a5f058a58cfbd83df9.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_01d3b034a2d8d0b83c0aefa4faac6c3f28ce737f.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5d707d065ae152450f9def619ddc3dddb9089e88.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_1132b11429034d96d82c82dbfdb69e460ad8a564.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_4a5dbf601de5754c03a03a1a42395dc0766fb8ac.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_5a29b93cee012c79d4364502f1d90f947c73641d.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_01e8f0df0c54ce619e5b66441b3c96a5e18b05d6.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_1d498e418ebbf33bed58b4074d1edf3d9bdd07c5.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_4d7dc0f356b630179916f8fc2041b7f1402b46df.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_292454f2d82184ab0491ea0675750c6ec55d659c.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_c538dc4f65d02776875627cbd20a9c794d70b043.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_2d1f2d1e57095f756ddd11e8e9d4f6f253e3ffa3.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_16f94f5c65c37624f5458c165daf83517d9e3c81.hip
-fmha_bwd_d256_fp16_group_b16x64x256x16x256x16x32x256x256_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2703018e71d57d3266fc35e2e18a78faa3dd52ce.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_ce5064e27ba427cb951f7e1b01328b0beb6b2b7c.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_aec87e65afa93e84d7a947c52f291c1c7360033c.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_30f0200092b0e18d57a9f5e512d565f1c0229436.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_61896aa9e4e4d7e494c1755b1e77a08e0e264f8d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_487724686efd35731e5335efa949486c93ae26e3.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_3e61b019e1398a6a3c36143fb84b5ff22c9f4508.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_94a94d145e575747c8956ac703810582c819e2e8.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_4d3b1ae63e127b6e6afe39e354d4995afc5faeaf.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_438e3565f4c720e6c9691b0d33c1392936e2e7ae.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_d3fce1e11aee2273620e75efe4aa0390fcde9ba5.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_1f0cad6ad5b172e51c569e84cd54a19b4eb0ed05.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d54b3731883a5f8393d60d27487f8d017aedd3f9.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_0efdaa9266a5a464009297dc59db92504f8bf1a3.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_99f8352674bd6bbe98944a1c0a769a4fc028a623.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_a5f2f0cef657ae5e333d65ae4ab20529a43cd7de.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_9eef1b54d5d3841f3fa6b84cca6c7ad33efa2d9f.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_92ba64cdf615c1be2865f027a293cb530fc07dc6.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_931cf8d05cfa45319f4e5bb49334d35a530bffcf.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_34807a8e90bf1cd839f32fd718afa6469c35a4fa.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_1a98bcbe900f8c141136d18c114b02fffbe8bca1.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_63f121a3c8928c10a2d86b487cd13fa995da670d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_8f607ee20c0d92b6dbd0338f139517fdcce98d0c.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_3a6b9566559ed2b1c85f2bea1c55e72c41dc47bd.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_45f4363f50af1e7ccd24751d5f5b181bf32c604f.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_22a07ecf1a59f72ec6bef3e970d7f33cf54c5f44.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_3400f0af03743dce328486f8fc805dd30bd6da31.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_9b841b7cf5da31f0c30ec42c91cc8d5bd3fedd03.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0e3f4cd28a4c06cc109f6a0798a77844bcc750b7.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_a103cd47156a98ad2cf2c325ea00df3f1d67fb72.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_bd37f4f7914805a97d5073f1ebf8a8b8c2648d31.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_030a759dcc92028b4c6f317fc230b98cb929e806.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8d79fe8a600c3b4e0ec9aa510f8036ba2b608985.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_10ceed95b0a0a01f844678717c88e0426fb503fd.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_90b17d8cba28cceddb3ef907df878aeef0762d15.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_b5ac596c636df55e81293228cbc53dcbb3024e5a.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_e68a9e05debd456a9975953f7b0d510e7a0f6978.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_50f915b4d9bd18a3c25a85917392ea4a5e88b349.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_dd67d442001d2b167e70e8730abde4d4461b8569.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_4160f6b6d0869740a5a411abd80108f729f810eb.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_40357c5e9739eae136a7abf92bc38d3ac94753f8.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b4ec377c44ac18527ca6a01bc3b146706a6e1e09.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_02d88a03cd3966dd0cff550065f58c3ffecfff6c.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_093834d4d3fe76e1745e4482c6b51b550c6f3dfc.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3cf45927b6d931e31e2209685d787efa28eed8ba.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_634d530731c7ade2c7beecfd1bbbca8583032217.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_311731442b756308c0a869f21b7b8b103aa613e8.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_2ae344010d49f7f9a6caab2cb84be7f87d2d96bf.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ae239476d61f48379754b97f29d7a285cc3192de.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_6e7e1d245baabe2f6293e3d85318f9936b333500.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_f6566441ac3074578cfe45758ba0583c0da0a5ab.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_de26a187c4db06115072a5132e1166b5b03368b0.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_71dcbe9f481c92215f3b636bc0e86ce8f65e6472.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_9a20fa19d8d30654602e363806f559113218d66d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_1e22f2d99804198c61251b4629a3f18ed3dcd42e.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_38abcbeaa4d33d3150f2b0238bb62ebbfe960980.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_d0863830fc5d43dc6d6400280e892bb7de2892d4.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0cee6b9427c164d78994150305a47f73954a67c0.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_04caeecbc01667ec6f5599358a0a20423aa9a00b.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_33099fcfc218ffdf69edb4f2f0e46121bea9fafc.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3dba3cd44f78c950fe7ceaa5f0629dfc607b30f1.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_c2f04447e6a94c94a2315454e71d7d607a9fd0f8.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_8e2d5f979fc4fbd0991581a020a414f9c8656ae2.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_1241814f76107d74ed069ecec99a248676487eee.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bd28203f47b6a48e9b66302cf8312f3796ca500c.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_3d289100991d4c8c362f64c8f6c4ba395c2f3495.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_7c23dde1a386436e9864c8fa5f1706c0d2fbfd0d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_bd8bf7c572c1984ca3061062cf3c31d993f6762d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_164a947a6c2ba83a5b1cb7074aee0bdac6c9c64e.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_9b062dd633645772e4f2caffd111af73184f7657.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_abf6c6412f9853855b74a96e862935ddef66f763.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_aebb2441e6cc1ccba4a391566e547402bcf7ced2.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_66968bbf7e210911fcb95ba90c79837230ab1ce3.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_be1e1533fc37b41838bd37edc2b6d2f2e76ae1c6.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_3a2280997eb6f1d091094fc54cecf42b7c9c3a2d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_4b4c03c916393d6be7c5181369ebcef949eaa763.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4ff20bafbf156fe8fb80bdd84a5d2f3a4a944c1a.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_be4dd90ccb2f258029d0156cf23f940b694cf08d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_e334e691714f0b99773c2ac515ed82de0f387065.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_62eb2f81e73d65fddce7ff43c397da6529317607.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_285e61dad8f63fb973cb2eb899c959e400622652.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_a2ef5d30a2318ae06430d17f84878800c4ca7364.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_54548ad36fb92d0963893146c8db20f53cbf0c8f.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_3967a8807c9451b09227c0f685c18aafeb062fd2.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_94f6f9dee9f0c3825d91f4d320a5280070e60ee7.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_59d366421e0b51c90fa53c366d47ed8d51b3a329.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_dd35634440edb25cb095800b882c70aaceca1dbb.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_0628931bf5cc1daa6e106cf60bb21fa1aac6b1df.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_ae4e7253ad4873576052ec0a9400597bb7975753.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f6f102a388ffb05c690a20a29cfe0b35a35eed61.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_235bf652702c2976551778b9159e09188575c63c.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_2a45129fc4995abcb8f880692f11c6186fc01641.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ff453e3bdc9752cb7b81f7cc3056325a8b9a8ad4.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_c08095341ca7e3a1debeb780c1878e351692bee2.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_5de27c4081377f59363c2bf2ea8624217566d2d3.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_8c4688cbd23727dd0ea9a36fb977b31aeae98d65.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4666db0ff7b035e54f2c0e59acedc2131b722a55.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_783ec08544591a22f59dc12f169b7327b4185a1a.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_131691f01cc7f29affb88152dd48c7a484315dcd.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_dcf815ef540060cc7ed43e1c57a28e1d080c5621.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_d7adde8780b39f1364c572a19c3bfb19417678e3.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_cf5c6c0bfaf98f6e655fc443246b81fcc730fe97.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_b18a615e66d7cd739ce35412811359a03cb23a8e.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_0fbddf533661642d84bf5a16149692d5a892182a.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_fde12cd366d6850ce26afce98e5076b695b4875b.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_df0b2bcba57e77d975ec5304fc50cbd09cddf4bb.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_cbe5a98163e878c7697e554758ebd0597c2c1760.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_3cb0cee09d633b6f70febbba63a1e090522cfb4a.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_eeb0e96b759e18cf703cfab0cda1385726f6e0a1.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_4601680af41c8738089ff377147e0547dcad114d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_6f3d098f8bb63133924aab70d26a6ed64018c13b.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_7d08373ace7087bdaca4ce8b0bc329f553f88d77.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4ec2075f394acfb14fae7b1ef4304fd9b654ba0d.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_8a1fd28acfe85b3adac859c4bbffa4d28fe634fe.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_7bb7b63e8a4c1df4eac4d978e166867195bd6e53.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_80fb694fce7b4c3c459fca43c89c6002fbfdaef5.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_86513d6e065a44bcb0c789eed1e7e5456e800ab6.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_31222e158484773d2257f4a31e3dfbdb68336a8e.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_b20c6252863a73341b0010191fad4c834860f884.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_70cf755f1485c065222be4daab84283a9c3d0eb7.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_0c8a0bb89a6f05289c0405df5126fa0cc16252e7.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_88ac7f6cbdfca2e397bcb86af4216e87166601c7.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_db8f0bd93b352d28c5b6d78f4332026993f0bea4.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_96c5e79f54b71677124f555b0ae4bfd27248d099.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0b532fcf26f90c82a792cde7943634f667c1d033.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_b6b17ae67adee9e56a022cd2a5514fb9c4e99920.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_fa62a97675719c2e8e9bb97361b92ff1c7b9d2ef.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_173c44dd85077e6b12dd06fdcf6b11ba349e1866.hip
-fmha_bwd_d32_bf16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f861d8693f82d22e2c5b1abbcbae5f30f4433e5e.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_970073c70133ff2ee4737f803a0ac43801c47242.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_5aba1183efe205af38e79a1b2dccea5fa515d02e.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_322a86568f89a5a5a165cfffbae9ca6949f2477e.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_321500dd4c41e4d68834814a48a639f5ca36a2fb.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4f4a5d56721bb1a1332a65882132a8c5763932ec.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_44d82b58fdc3e5b7a7c20490ce7f5acce4e6ec79.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_678a4a8210a972bb2ed89d6ac754fb79438ab2da.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_085722b43cde5f37242edb071f639da7c4a0bd48.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_417b1cb14b67dc82f614831550f7deb0895bd7e4.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_7ec04763d635c5bc3e810737b5d948c59f117d5a.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_7524904ac5a2040c7ea72aef5942212f291a21bf.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6979ef43adffdb62100270a62706fb811963925a.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_5be9ed84ad9be1627db7a66af9370679816c0897.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_ee239db5a67c23a383590a651f0d8a0be43a13c7.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_69214eb450c3b249017480efb8d092b0edad6dc3.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0c32a2d9701e23dd930119c4ee8089042b5b0ac5.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_97246460c21bc66c0f13936d27477a9fca1c44d1.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_078b96ad691a85eebd18586db0b62b8911016d9c.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_7ee953cb24e28bcdc8f05783894b23cbf83bdf35.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_921f789d619db6f225e8e9d646e93bbc9dc1a669.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_28e4d2c757e4b8c366a2c320360e21ff0ef671a8.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_da6afccdee4107507a64323e17bf12c46da2b92a.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_38e12dad9e3bafe177ed3c27c833825813e18fc3.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_151a4425b411596c46c7032f6b83d3152a0e0cd4.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d1d3eacc320104100bce46235fe656e5a8223c66.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_a71305f191f06cd53b7563971c706e8b71b19e2f.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_482e34930d11ff493007b1613993e01acc1af78d.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_342d29c85070f488a14b1915f948e5fd69019c99.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_b0f555b74ed36f1bef8f47880b3edc6760f27788.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_c42ab428503e8f8bfa78c8cb8d9afad9f5185118.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_cae6c7efbfc831e2bcfc8c1efa1a486c02627cbf.hip
-fmha_bwd_d32_bf16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bba10ecb79ede07324e1198a71a95ff26e9eb235.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_48ae3af78583258c4b13c11a442022e0e058bb85.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_82048cf91270631f98ac37dc488a1fb2e00ce004.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_fb4c5f8fecfbbe16e6648becb3b5ca89fa3d8a94.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_6abeb7b50ae6a1fc62535b9a1dabbde6f177a9d0.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_8a824621a50cdc3cbadc4b1f9ef18e1325385082.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_f69548d6cced86c21c09c6475237a0cb926df0ed.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_325fbcb9e503e68fafea08abf86a4951f440850f.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_964f916d3484295b5918e2e4c22c5529588a5662.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_df645b3888dc8d1df50c47c0d75822eebd3eb019.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_75a310a6eb86e3e8baac7a930c3ffbef372942b3.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_458d708d13577f2b92e6d5adfe952a87e0cf7be5.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_15fe3e8f4add16a088fe44458353fa7c0c4f9658.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_0d0e0147a92061d32608a34e7b47bd534eb787fa.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_4e15e4f16de26068cba30ef12fc29332d45e460e.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_e2bf6805a489739abb77c13173d57723e9304afa.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_15cf7068183421b141ed5d6e7fe902d06b6492a1.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_703246f1f53a988cf252eff88bdf814bd382d3ac.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_381b29d9888365bff0f109d897b508eebfd8a61f.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_f2da112b1e07c44fc8a7f19368da203f6935049c.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_1886d4bf54b3a4a9e093360998b2059b3c03d072.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b60a4e87a7aabfe3c1ce02b408522f3ec862e3d7.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_a62a2ab489839ea1a1bfd1b24e54a3c232ed934f.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_36a0a960541bd8a2dc6741579de685b7c0a5f6d7.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_1f6bc5faf18be193212217788d476ce6fd384bfb.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_13f747525ad31e76c88774fb2208e470da9c2310.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_71b6100efe30d836dab557ea4ac54c4b9d35c6aa.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_62ab710e4acc711430745e05e036dd6a4d6bcdca.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7597ce4d2e5264bdeda47487d5bdb55a014c6616.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_ec7ec8d547ee9713aa3b5b667f22cdcaa8f62b2d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_4fe530cbf6363a8f08a94728e45e88ecde299e7b.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_661ffaf653085dd7f122d603bb3ba4b001e5f3c0.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_345ea796c8d97bfe3b7c9663bf15e2e5e7696235.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_802b21f9588d72c3c3e3b9a3b269f19c484d5aa4.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_9ae866c7db36286876818bfb718ac35204fa3843.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_faf56e45b2240515e97fc1bfd552eb03b6de5094.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_fffbfcac254e33926131a71905e93f9cc0aef89e.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_919ae177b7a793fa352c4f6bb8e4175f3064d814.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_ac9382cf8bb56ffd962c99329bf67da992f8810d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_cb1deea4f4fab0db31d46a91228601f0c272d6e6.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_144f19363ef26efd36f0436cfa9f84f181a8824c.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0392491c5a6dfc742c2be483419a40f6a7a7ea56.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_cabb7b12cdd9b8b522af577e13232b2459dbd38d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_caede7a18f3e3d5e24f6c70392413a2cda16ac15.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b9d00ab8373747a5c6b9d2f8dd50ceb14db4163c.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_ce909cb5f96a4884caa0d2eb8c5e6bc7fa352797.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_1037f1bc50c4a65dac09ba56b701256b701c4322.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_aafe891dad43815e635f81225705ff944f990d75.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7afd1a756247b15b078d15a39e350a07c22982da.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_3e839660557dee9d5bcda9b56940ce23236c5f6d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_fd26e43ca652e6f58ff48c356165aa4349833b55.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_b3486244e0b7d6dbcaa1951e8b8883ce441c3f99.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_90da0d469cca5c8481504148468460c85a15c559.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_714c5369aa848021e020d874289e3ae4e0f74d77.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_cc54b107e1b557ea36b5cbaf7fe3dfce05415c86.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_00042c36bc588e60a7c8a9ba297a8a25d8ac0660.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_65794d9c185b21f59274ac5d4db10a7abc0be968.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_faf686067fa433cea5e95dd523846dc881eff635.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_39d3071347a0c98f3221104036f477aa13bffa4d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_e76879f8ff4796f48ad87ff8003f4f6e6adca9a0.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4377ac04be3a6cbdbfbe57612a469412812fb5b5.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_c1b76bc7a17f573c0d52c07ae9ff4302662ae61f.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_d713fe25dc90b3511fc259cebf463376dcb55d84.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_89a3327da9a3411ff1cddc67eb647083cd947a92.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6a7eb3d86aa385f9ecffbc5ba10489e56856f918.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_4d65e58c9f147498ed04dd51fe1393770603a6d3.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_a5c0109313de1f6245d2a80f8539485b849e9d55.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_e73a776ae4ba68c23acab1a5a6381684051738ab.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_a225c4f1f3c7b271957768bb9235131c67afb48a.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c9530e20038eb40c49bc8b045be0cf4e7e6b4eac.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_f51f1a11f778d99a00aa5959a3e58a41fcbfb1e3.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_de36bc309877917a18fd21acb30563c7e2f233c1.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_4b45948f2795293e72530b02669c4f549608ea7f.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_05f794c7023cbb7e35f1fd1ae45bd2377bfbc520.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_82f1d7e1a93bf2fa80c409e6827ea88af56c44f0.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_4baf664bfdf070362bcc91af77d1bc406f744351.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_80efc341089a50ed5669b3c86f6ddd9b124d1442.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_e465193d97d43237c22c04478ca5833011d8dc8b.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_915b75db795dbef037b14b003ee073665fe35d3e.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_fb5bb49928ce5515d7b297d5eadd4ec70a22d60b.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_349241529745bf138552f49d9a93db418663ad65.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_c4de1bc135191f3c2aff740f4c6bb7e98da42f84.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_4ce03571f1d2779bdeaf0a6a2d617e236d191c11.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_ea077e68dbc1bed2dd20a5f4dd35e0cad6330ee4.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_c56aa150611b0d4800470c1493dc907082a5c23f.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7e9519dd0d0f940fd5efd61bd32df7528ba7e3fc.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_47548aa042c69bb9c59a8bf706b44028aaa41830.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_3dff884e176ec7cff86d17c6afe1ddaa4dd6007d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_2081430c92864c29bb9f409e7c27caee1de00749.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3d1cea88a2277b87d405025ba256272a1720f88d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_a55c7dd576e5b1061c059e5e99aeedf4389e2d25.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_8c074afcf33e3f3534ac3577484237fcfd2ca48e.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e618fb4e529104fc90069c8779ce5463460bd516.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_44462715ed5f192532760d6f4c66ff9d4e20e254.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_e1d85ad2c9d197f501267fe0804e6985802fbd18.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_78663faeb0425f45e8a0da0f7b1a5ddbee5e07e7.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_15e8e1ab8c63db96843054bb7a98d708ae6a9c44.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_629e0b97b3fece7c12504f4c8f1860d611b57269.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_29c9e5384809b21f39e78bb2e43af345a9a21d19.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_14f77aeeafe4b28f314fde5ebccfd2a554872781.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_09d76cca48b71dbcc9bd96734787209fee4c9a74.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a4980becb0d3149fee575bad1fc3b463d08aabf5.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_55bf8444c1c26b91fd490c7216f4d0f8aa0a1f1a.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_e4d9a2396ceccdadab24602f30e9070901a76dc7.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_14fea611f3c253aebf726af3e5fdb7e63e18e13a.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_587fc33d02b1932235b8d152e57559060211d591.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_680e81c3700f130df142c9a37a368944ca548721.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_62048a8ae1c0096f3372b0114c15edbe813425fd.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_9b4dcde1ae3446b825dea739d4295c1d1ec5c4be.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_ede81dbc4cb208ef6e684c76ba1eb451d37fe10c.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_59901147b7188212b8d8feea15831a11425fe4b3.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_c9ad71883a19b522486706d3705700c012a6fc19.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ad82071cc074fd30437f6158b5eb2c6df1f8c587.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_bd3daa5f99b4522d932334924347353ce2854821.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_b72a804bb3c99830653d41ac0bd49943c801b89a.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_572e68bd619e118292768f0925ccf92cbfa68415.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_ee1a43f2210a8d1e5623411c95c33424cee5e747.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a93a03b33305b33055273711ab31a5b8d8298d5d.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_c3cfaf0d53869c373f6d0ec821b008dbb819141a.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_b2af5f5b5ee3ae964824a3e9c7bbeb5bb39c557c.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_56964a17f902257aca9d08c736516a2c67d9a0e9.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f9824fb32933b27501ae8a7f43f460a2dda6a814.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_4118e3ab290263ed2576feaf22a1944bf2ddcb7a.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_ce5ad502dd40353312d561e9f40aa478c16ef5b1.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6d07bf9c05e41dcf2416e05dab4bdde17158db76.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_b1c5d55d47d6038e9162d32ac968ff58c0942938.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_30c8e4d5c761fda50e010da779e8e4730051d403.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_193699a5daa14ca2def07489e0b563149bc403f8.hip
-fmha_bwd_d32_fp16_batch_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c0342686e4efd26413c6719782ed13603479c4e0.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_fb79e1f9231692d736dbada062ed6821f34927bf.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_5f3c3bed2b584ea2031debf9f953f5f8f7012171.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_319df310195191895005b30151da8c1afab6c82f.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_6af23d1460abfe875e71f7911697c42fef0f41c5.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_cde0582e1aef74f9209de638b553ec0671476258.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_5052b2318dbb78b1a82ef03666a35a623f44481b.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_2543da478310245e19e6c6a0d9ed7ad99540b3bc.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_20f7ea0aabd069362ba4bbd66623cea5b6e1a6bd.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_da74887afedbd67928fe4d596709f9ff92530611.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_155c3549d067464d186a99b8205317cc000d4898.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_089a347aef8a920e3b59d5ffe71fc5bfe002609c.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b0dd965d5d9080ed5c6a04b7eea9890f3a264f20.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_2db33b5442d2e0948762b1f2147a321a9d6907be.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_1cc459e57bfed5ec7f40ea4a4dd9f72f3ad7a709.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_02ff94e3c787a7b06ffc90c25777fa74f225e32c.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_86309c036d96367939ccc3e8922595ac35a3e179.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_8b92990df507e82f96eeb7aa3ec00c01437566fb.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_26835ba70606c769e56d19dbfe74061361aa855e.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_dc1a7f9b1afeba6690fdc0d0d1755ea89c805573.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_815918206483d2ae04a45aa67d69dfb986587214.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e1c1a31a1d8556cbe0b6ea76faacc78855108539.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_54b6e18b10d529eb6b32d7c19c59eaefc7184376.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_a622fa57764ec746e02f6d4bd4846b48c722b807.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_3b5b3c218e4a7b459e54080e24c5b730221eac02.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_85fdde4b25e2fc8cbdd46c2850c19eac8d9af8f6.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_a4b7f10440331a8a88ff93ba253217c2832bcf9e.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_4e9a933b916285d9580a76df543cfafc88a536cb.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fe8b8c3525fe86a20a2d6c69585f3e36c16caabd.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_7d12e9cb599d24631c082e3cf65d2c58b6d4d44f.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_8e812705ae3e452810794fa7caceef2ef6066dfb.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_7cdc419d4248dfdeeab1f0980aec35fa134e52e0.hip
-fmha_bwd_d32_fp16_group_b32x128x32x32x32x32x64x32x32_r1x4x1_r4x1x1_r2x2x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a046e888e3836b0bd3c49fec8e1872e880798f0c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_8278845045d68027dcf3bf867ecde2fb12ec51d3.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_18a4d71b31c451a50df7996e3db864bc3c3882ed.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_5c36fc744dfb0d985c9113175e76c7ec1c935054.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_b779cc0b0380e1e6a2b51fc6216fdd72215b882b.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_19af6a7f9e5020e8d0f0ca0f6258001f6ce592c1.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_459ea3713aef9b916e1b38a882a45012930924d3.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_977137b371df841993c8d0584be7d83aca6add78.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_7497eca4d1a18306b406b367653622a8d64095bf.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_9bf235679af1ca03a6e601b4cf6cd0416d1c9091.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_7177f939ac3dae8749cbf4232dcf04d2cf63b48f.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_1847fef2c06ea581b0ab31af1cb0556c572696ad.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b5bccc85f74f54a2ceb17fe3040b04fe306c53f9.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_f7aa9c39b06e55bf4bc9f9a2a0fb075c9d4e69ce.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_a78fecb9725ceb4bcf2aa037d43bc43efeb1c3fd.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_f93bf815b520a9d9e17b43bf9d7fb870751b6225.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b24f91dec2029b25d0d96962528410df55a468ed.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_00a2adbe938d458d51ca5fc4020667a215b672a4.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_036887daf6cc092e7422a17882488e59cecfb643.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_1a96f0ac76f117e66eba97cb990c2350561ec2ab.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_0c3b2ec99fa7b09c7f78dcc3142a661d686044ac.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4e760de14b71a41882ec4a2c7362565af36d1a5d.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_94aa519eb57e5797125728492d9330f5c0f0670a.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_6bad2ed9f91bc1efd89ea66cd5c775fa140cf931.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_9b73c92a13757877f34bd8a13c6fb29b60999020.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_1dc6e599144a093203fd7f92ac6d3c2cd7180d49.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_7e6129eead18d13a4a6cb9550384fddabc7a2a16.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_04f39b453505f68a5091f68b1c3de48369d1e7ea.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c5b440ca9a5196ee1e72c878c87d96934e9273c8.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_cb4576e8ea5d59d7663f3760009a00a19e1b0667.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_44690e48f30657b0fcfa26fb3b9af3ef76e792e3.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_b872f9e6ebe330cc1818ea82b53acec79a2f672c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0fcb7492feb79e27e0bda73e57ef7dab410e2bb6.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_7a242e5953f44316b6a4f6587ec26283ed6cbcae.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_2184fba2eec5899bb40d49d4508196e6be1ec1b1.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_06b74acd9abfbd1c4ec2f4c718eeb92a0bca7bab.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_ce5c161b725becf059fb4439c668edd454ac77d1.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_addb6a14043c5a4df0f5042b3770b40c4e90795c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_7ddd621da88c57798db1e689b93b692b6519ff96.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_b0544a38dfdf4d81dc95894387845f48435e299a.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_11ff174ff2175e9ec22ac3a0fa59dd7713b79643.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a3f9c236d24b30bc9c3fad90cfd6eb00da835de2.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_515128c6978449b33ce0c35b02a9e9aaad65ef7a.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_0b3153af7bcdba33115a0d31f121fd76be2ffbcc.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d3a2edf232786d458e2125f8dfeda8847f842afa.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_a7f7553a7d2f6d42fe695cdc64423c85223af440.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_a9b50c6ebb27986ce5b378d8c39315eb9cb91dea.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_2f55a23a0f24ff7062a4c286944f25d2db3e20a4.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_0be8cf70c6be969ecfca675782c860b5b75ac089.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_1e9130607a2d24cb0662a47e9cf12c6602143838.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_cee81ab2e2678816c7b516d2d4c50e8cb5874c68.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_c5fef330a975002ed15670e8e7b26a10376d3cb7.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_0c9bd38b8f9009d932ec49204fdea39a52885246.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_82c932e6eaaf44861c794539d9caf8b50192fc44.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_4568af1b2f104664fd05d21ad789aed39ecfa42b.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_d9c23b7f8fcc4e4f4c81f5f00cfd345b98df2e0f.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_de7eb562a7eff31d589e12945d80233aac202ae2.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a92b43d374642df991edef1f6036dc898bf77cf8.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_014c209d5cfc6b965bfd78c64bf132c0154e32be.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_1687ddf65ce4ed2997583e20fee9f201e86633b3.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fc5841a729099340d608e31023acbeaeade3e886.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_9cc3ef3d3b36f52089548e9dce522b0448e2c26a.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_8efb5fc2ace6839eac741c5e6616665845f43566.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_ef5421703cbfa63a58ec02701e245d479a1fbfc1.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b50e6df20a2426abd3d2ff2262a37c009196024c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_a094599fb5caf5e7aba728cd4713a8d0c6368a46.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_21e235e31d6955393ac8e825bd69ead70687b7c8.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_289071756e7d0582eb61ce6483fa3c988d2e10b5.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_1899e28aff2fb168cdc3af7132dd7fd09c2e1ced.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2e30f50071113dc4ab59468d568ac9deb06b0342.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_bdab172627718278a71a93e3737ef08ad9259a4f.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_77200e875e0ef160b311c7de450c137772312d0d.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_cb1b91c16e0255fe7a0a85638b98d94634e143a9.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_877e33463b3bf1853c6d2d2009af8d27bf88abbe.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_92e53359c69bbe4d7405d45261a8a62008eb7d06.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_7764814a0de7702f0b7b5ce9dede6440603f4853.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_05dfe927fd64a564c5fad537fb7c41ee9c94c2c0.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_78f7e2a2c08cd87702793f91b6935cbe4c22be55.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_d4605b2ad3e3753c5f255678abc1690b949c5abc.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_037c6c80fcec3eb8b0bef50ad6af6d27bf5447f5.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fd9cd1305633b62b68fb8474ce021f639f8492e7.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_d2f4b869ff23874b6bde0aab68c419108b7e69f4.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_6ff58a5186d69efd6062f3717bd315394ea6592b.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_8021fa266c77e6b5bd1af2a9c22c686e5a6eac78.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_de5359f0fba3da9dfed06ddbea8fe2a33a9cf40c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fe72cdd69944d2d765478d4aed13066a02b76f6d.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_6a7b6781ffff9a42beebb4d73f0d15461ddd4479.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_28f7634d29bef11fd466b452a46b0612f38c949b.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_66f651d3415562206c1049b172261fddba01ea6c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_32438250078ba2a47345ec4955dafb4e4de78a25.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_5ea53f7c6370845fa94aa9b395c52fd1900b62de.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_d50ac8e8a03f8e7ec2c6e993dd39f09f465dab57.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e2b629c37cf94134693ce455b8c88b72a39df7fe.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_157b89d8d625b8244b5cceaa4d3e5fc5a09c8989.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_5789f267d34c9961ced63ad07ffea2c6d2911415.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_38010c9bf7341588f071f889b7a0b4dcc4e7a14c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3d55cb42b0096a8ae338ce100f86e378aa1a04c9.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_11e7df31541c3aa919e9825ad7dc4432f9a03c0c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_d7145383e39dec0e346b5094401acf85ef3c2075.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_04c363e11d202c6d2f4bb753661c5a2043edc0ad.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_8fb33fc20f2e85e915f1b1529ae87981dfcaf86d.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_97851d5ecbf02f8af623988b1a39c0b91e51533a.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_9163ae070075f26926a86d39e15c27e6edb1f1cf.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_9ab73ea77ec20ea3bfaf995dacf93a6960ecdca0.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_21828c7d3f5574690f12f841c27f025206e6165b.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_dc08afbff5def8bcb4e823657ce01f57c9dc77c9.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_875b08ca602fe48840c72cd61798acb98540fcd6.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_216806a4598c885e517e664fc8280c59ec3cbf11.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f90410c26d7649e21e2ae5e32e7af89d84d2ea70.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_a3339150d8bf9d073827738527f6cbe15b854607.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_7a0ab620e6d62259a559e329460e46e6e3f7c3f9.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_7a2e032f6500fbc5468183415b6dd1d3e43f0bee.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_71a2d046629a4b65c90d0e18d061c4984062f844.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_04ffca078cfab8bc6c4ccd1cc8994a1bb4a88ea7.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_836a308c2d2afd6e0dfbfda61984b631c4ccffc6.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_62ba7a5a0f3a714eb5f9f2af20f7bfbc82a30350.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_e9b04e6d5527ba0b8089ba8bdd264e2d5759338b.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ce5b5932f6df9a194ceb0d69220fba9596528eec.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_cb3d5273945c5d40cc05c2660af2df1fb7a15f3c.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_5ace1c9b00f160a17355d4583d49c47887ac33c8.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_054fda16133a0d25077967b05425f9128e1fe1a5.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7adf69b51f0a8cc9ae7e250e60df38758230fe4f.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_cd757a8bbeabd16a44d149ab188430f6d79ddcaf.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_a5fa94bb32a80e81886b711ebfcf2df5f5405866.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a02f152e9184af0b3d77082d8bdf519dbbfceb2d.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_cf73e1fc0015094861ca0c1c81bacdbe0c5b8f37.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_a9df9ac4ee78e5f4d5bd0567e58a7090907c61e1.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_92121fd448b4640a17e1a7fe73bb7b58714c0afb.hip
-fmha_bwd_d64_bf16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2c9756060ac0e73dbcfc58a9222a78f0283cd029.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_354121d3bad1d448bd413718fa096f54faa12e95.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_d4c9f975891087e6eed6393629b41155deafc509.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_9bcc791049e3ff9ebc1a9085d2d20efcc2f99b71.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_79d0b8053ddf99a4d4447656d733c2da026b3a7c.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6f8788c537cbf6833c58a6ca15c0a36de33c9fbd.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_5fa19223cf296d7fd10e15e2571e63c84a80fbb1.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_4dde56efe17f4fd36a11cc959320a5e43f1dc232.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_fabdc143c29d5ca50ab1e96a814bda6d05b0d5d2.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c977735a36c325706bd19a12df66ed0839b032b1.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_7872c45ba170f2782c4b5b75cfc78ac79a4cf157.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_7c4710e8f4e27fae4ae079f1667c3a1879cb6da8.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_09e50367b62bb09071e28b44235a7c112645a706.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_1be43f8b629e7039f57b95866d5777273377470d.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_d0de618ff3ea9f67b90f2227fb7fcc74ea34183d.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_865eb90b1a2d64acc0f6fbe1d807c501fd4be3cd.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_bb35c86443cc9ea38c06ebc0656306483c95ef67.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_ec171210efd217c07d357fcf42e5372ad7e9abab.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_bd80a1774d8b7d8bee4e8663392b97cda11dcbf5.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_b19f05f6848403480ba41d37cdbf44ccca1b1f8d.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_e639a1e84faa98477b05df71d363b9ff0f9b2760.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a388a284f45f711d82a6ed87036d87cef1872eb1.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_26ea90eb5a527434c1740933a1d2dd863eccf14c.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_e16edb824cecf459a8ec51b8dc74b1e06369aceb.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_843e7888cba5f463d19fcb71aaaab25dc3d2c09d.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_4f6243c6850c0a2d2b7bf1476e12f95f187257b6.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_89617bdea526d12d6a33ed42b9b0018c0b173722.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_2b4050988e5790a28dbe10b4c20e14f10f6cf85c.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_44cc95831c347212021c0bab7b43acd7daabce42.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_ece60111633db08f765b3c7cd5cd768cbd030255.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_79a7dce707954e765d97cb22e57d9bd6168860d9.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_761bde840c0c8149b24a8f6f264e963c4e9e8ceb.hip
-fmha_bwd_d64_bf16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_609616f72bf16a060fa50091ac139ddc06bf9d88.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp.hip -> fmha_ck_autogen_ca1992a2634cd6674076611be54197c715ad8271.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi.hip -> fmha_ck_autogen_2f0247e301a7b076b6ec8a778c3b47e330638963.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_deterministic.hip -> fmha_ck_autogen_55b14cf2998a61611d1de2594e926fcdc378999c.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16.hip -> fmha_ck_autogen_21411df58165946bf02942b597d94de7dd856987.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b3063d06723ac70c5f8802ab49c5c35e1debf56e.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask.hip -> fmha_ck_autogen_4052ca6a3ec02f6559e4bbf1edde42ad2d127c26.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_deterministic.hip -> fmha_ck_autogen_d41cd6b60a97e7071518cbd1a63abb8b910df024.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_e75d492ac3a6ab75648056bcf26250a4aa929cfd.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_474fe2d739eca8c93fdcb2c105d4154cee6ca1c1.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_deterministic.hip -> fmha_ck_autogen_2c0bda0feaade2b554d648d72f219ac9c389bf09.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16.hip -> fmha_ck_autogen_2122c973581930ab7a4ebc90b3bf1cdaa229a87f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a20c91b2f11bb7e5058ca7935b0bda4f5558a9dc.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask.hip -> fmha_ck_autogen_9990e6ad243a48b84304b5cad0c663c0802aedfd.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_deterministic.hip -> fmha_ck_autogen_7264e378e1ea1d4dd97f6949d66f3492883b663e.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16.hip -> fmha_ck_autogen_7878e2a4d3b96a552e03d1ffc33debfd50c9f7f1.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fc1eb85a00017efdc610e4259d2abe935b85304f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps.hip -> fmha_ck_autogen_cbf3e4d4d4837a0cb33b78c4f2767b1d93da0850.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi.hip -> fmha_ck_autogen_5f8925f929a5b26f3544ca31938aa75b3c59d34d.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_deterministic.hip -> fmha_ck_autogen_8004763f674dfb3f14b66dfdeb2a046e413ce2cb.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16.hip -> fmha_ck_autogen_0878b9aa31429d23a93cd953cc6a2fc5f43d0d3a.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b5ba2e73df35f6e0f7317303823fde92a42b1a35.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask.hip -> fmha_ck_autogen_d34fcb56caa8f80404789fba0ffac447483a4d84.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_deterministic.hip -> fmha_ck_autogen_cb1a0ce432c27f4cfa51731c3ef181bf60c8a727.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_efb9e7d9af47cdf79f15f674f8976c05f08b0ce8.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_357f7e626135cc9176a295f3d1f336a7c3852688.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_deterministic.hip -> fmha_ck_autogen_22c142d869ef940ca876c93033ad53b576ed34f2.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16.hip -> fmha_ck_autogen_1621507cf219fe608715d4e5bb6e5764022e2d61.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_dropout_wg16_deterministic.hip -> fmha_ck_autogen_a25e2aed617e1ff31f93ae7e054313ee0dceee97.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask.hip -> fmha_ck_autogen_7ec038393ec329a894aee9bbac078a40f57a4684.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_deterministic.hip -> fmha_ck_autogen_15dc02ea7e0908cf0bd48034f5a49debfaa36219.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16.hip -> fmha_ck_autogen_758b211174da0f398b2a093e7389905b4f9c4060.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_ps_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_548b347672451e8391388a400d016803f4c4cf8d.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk.hip -> fmha_ck_autogen_ae7899b1ef159ecbf01f27014601eb79b31b49b3.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi.hip -> fmha_ck_autogen_b04f14f829eff73afaa57a875f74ebd1e6860979.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_deterministic.hip -> fmha_ck_autogen_2ad492377add5c8f6d0d2dbf9ee9e4338bbd9f1f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16.hip -> fmha_ck_autogen_7f6ccdb3c2d595fffd05bc5e6417b157276547fb.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_69cbe8eca7e3510f5caa7f13419cfbefbf031754.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask.hip -> fmha_ck_autogen_8bd7b8c63a51c8639b3cf27ad09d41ae47c480d3.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_deterministic.hip -> fmha_ck_autogen_f21596e8c608a795ff971aea8e199db9e72b65d7.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_1da23de9604b5d98fe02529075bad995954c12ca.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_49d4c005d723cdab9fbc307933c1257d114b539e.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_deterministic.hip -> fmha_ck_autogen_e2c9f955f227430c6224ebc347649386be7f01eb.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16.hip -> fmha_ck_autogen_290c484c2a366258941ee0051e139ea716a9de2f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_84cca7528c7d1bf49ba79625733ff0ae7522c096.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask.hip -> fmha_ck_autogen_f3d0166931e4406873d8f552a5d5b61fde2391a3.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_deterministic.hip -> fmha_ck_autogen_8046f566fa7188c92568b277354e8b06ad382544.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16.hip -> fmha_ck_autogen_12d60c8abecb3bc9b84b0ea7851628ab17d8b0b3.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_psk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_f50fa4ea674a590d0a817367ad9915a5fce20c51.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_0836d5dfc0f939ab9a4064b403339373caf35b56.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_de6683d175affaa5ff261ab8503f64172d8eba8b.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_beb9afccc15de7dfcb2e7d898abc0d61201de73e.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_e6e0ec1db1ea308e226f675e68e29b839e41b252.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_7c3d8ef4da515960bf40eb1feb04d21950ad5ae5.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_fcbe827108d252b2f5847fa8e132c9c3e56a90a0.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_7993fc08ac5c6ce7a2eceb1227f4e3718dc4cf5f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_06ae52ef937cc27c544e32025ea0dadb7fad982d.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_876a418fbe6183d0392b7a7d9986d067e323e2b9.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_b03ab68e33844f97aa58d463e00037bc11c50da0.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_8c7970957024de050748d3e31cef434f582d968b.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_add29e3e9828911a117dccaa5650e77805730d14.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_0e007c36231ccdae12f102eacca1f74b0711b9c6.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_765940baaaa2ae6ade43ef4c94a220eaa63702b0.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_c7af2bbfac25de2853be344b9f636226c1c0112d.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_b2f91e937b427ecc932c0cb0c90b2c2378db0be6.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv.hip -> fmha_ck_autogen_8da8285bd6182355e3164cdc5a983375cdf0a61d.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi.hip -> fmha_ck_autogen_a3ff8445ba691807caadd9f26e7eb90851875280.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_deterministic.hip -> fmha_ck_autogen_9c4fc7cda4b560040cec93f63021b529aa1ee3fd.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_4018b1fcee808b6cccd131418b6ae9e8bf900d8f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_88d52c5f70abb525b9c8aa8fc1cb3997c33ed67c.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask.hip -> fmha_ck_autogen_99e2f290b962f1617b0a9d4fd6d55c43e4439d6f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_25938733446b6c0dcd159719f08d04a9aa467967.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_76f884e9ca116ee47b446efe9fc770c178a858d5.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_42e2326066c91452335eac05f25a6311376bd9e5.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_deterministic.hip -> fmha_ck_autogen_24643917fc970c043d1c80d8d4b17ec92deeb8a1.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16.hip -> fmha_ck_autogen_d937609afa8e21a761dad6b01ff3f26346e450fc.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_59beb9cb4e161f9dcff79080149076488d436301.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask.hip -> fmha_ck_autogen_fd3558b4c7a667dbc365c4c2ceda646975408f51.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_deterministic.hip -> fmha_ck_autogen_dda8d021381083bc48b7fb1840729254dd8e5137.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16.hip -> fmha_ck_autogen_ed37ba962e0288e2840eb0925d016b5a7e3b3164.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5467aea26852aa9a9e3dae76b906005ddf6fbae1.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv.hip -> fmha_ck_autogen_76be322fc072ca19baa82707e260c6eba936ae19.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi.hip -> fmha_ck_autogen_c921a4790f982d48bcaf950123c699647afb739b.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_deterministic.hip -> fmha_ck_autogen_76674fc182dfa6329c73a354aa3adf458429444a.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_54402a22ceee3b665a3f24edb98b8398c35c6f5a.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ada016be2bd0e377fbe01fa7adb9bbb8febce100.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_6db86621d626722434f2ae9b7b8ab435a8dd8827.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_162b0dfbe3f615b1d164290799b2457437a0044b.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_628b28f65f19e7d1b22fb3b85b7cf3d09cd54ebc.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_031b12f9fd94e01aaff2c0da4f35f346822087e4.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_deterministic.hip -> fmha_ck_autogen_b9a742ceeb6736a2c8f9439d0b05e10d3e0c5c6f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16.hip -> fmha_ck_autogen_afccf699f593c828e11efc053b144044e45b32d6.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_fba36678d5047ded97ee7a7ba9feb9569afdb6ea.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask.hip -> fmha_ck_autogen_14baaaf1e90a075ab802c6e7d97c4b1605c8bd72.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_deterministic.hip -> fmha_ck_autogen_0237c76137df14fb808ade8bd6837045f2aaa5c9.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16.hip -> fmha_ck_autogen_c2a2856bf9a81544a30d535a13554e3a8107c476.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c2940fd05efd52bdf8a3f9aa4b78bde9b5809b34.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv.hip -> fmha_ck_autogen_d049a1b8f4c1c6d37973ce38593efda1de8ce0cd.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi.hip -> fmha_ck_autogen_f4b87f983a5e84582efa1663f84da76cf60b5f6f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_deterministic.hip -> fmha_ck_autogen_4db2e63cfebcf84043f79be0321708cd159c62b9.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_f25b87c435bc5d7d85d738f3fdf68947d79f5a77.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_540bd57333c6839ccf5cf2e928edb996bc60c371.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask.hip -> fmha_ck_autogen_9583148fd684a7e6a312127e023798278415bd27.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_bf9cdf86a7944cd690b0fcbbaec235863acd10bb.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_2da2b905c4ce32234c2af62328adae6b1f9217a8.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c4015f0d0a7a5173810f6f17c00065e03fc61a89.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_deterministic.hip -> fmha_ck_autogen_d773df9ccfc1ace90fe3afb5c00976deabedf6f8.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16.hip -> fmha_ck_autogen_d137b7b6e04e1caf43a62bd6788a75361cfa98f6.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_adaef10ff2c5d89530310bdf1d53a194f06a94ef.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask.hip -> fmha_ck_autogen_1be746990a2032f0363ad9f9112cc994983f4706.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_deterministic.hip -> fmha_ck_autogen_55bd9c4f1b7a0621c67f3e964d946ce22fb2fc80.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_4dc87b7d385e7b092e4706c464217b004fd8a6a4.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_pskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_91695dea4171747fb3cc6d910459f800608d07c1.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_c137c03bf161b2ec6a9a046fa49d7bbf80ae47b8.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_83080406598df6bd3102db70a554e496e29db96a.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_03a71615a088e972c998f9c7cb44566c268c5124.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_6214f820b39a8ba81e547a78ed19a909ac13221c.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_3e2557f206fd81d82a3b9d59113105040beb891f.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_461737a13e24009bf1a5a4b780175043a9f2e33e.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_83f6a1837a65df12b7c55d25ca28cc939c2a6328.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_c59a22c6efd8bb8815887325aa0b739e260cc754.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_6049c01db99fce654e9351e711b113cf7424550a.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_c9f28230817c9d9805c41dfcd4e834fe302e1df1.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_7728d5bec7941c9b6d5632bee8d67ed92b9c03ec.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_28f1ef32c4384ec26f3dc5e3af6a74fc8cebae92.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_594929c433b049a8cf949ff476309a8faf5c25fb.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_8441910c34830ad2459fb85c2c14af02da718fdc.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_230861e81e5acc523fa680534eed757b7b4a4e1d.hip
-fmha_bwd_d64_fp16_batch_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_c112c01d201c366bdd7acccf2e1b18b00f671153.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk.hip -> fmha_ck_autogen_6b638314efcc4f16aa4a6e58e6caf2fda1711519.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi.hip -> fmha_ck_autogen_c8f6461673882d636772ae4d26e78eabcb568f31.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_deterministic.hip -> fmha_ck_autogen_f93bc23b8a4f1e0fc5c5756c4e1c835bf59dea09.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16.hip -> fmha_ck_autogen_4356b3a2ff49f72b91a6b9c215df285f2798ad47.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_e1cc934ba7baab1a2eb062df1e4ee5066e9ffbc3.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask.hip -> fmha_ck_autogen_137fa6780d9e6bde10aec10a875c039fdbbc652e.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_deterministic.hip -> fmha_ck_autogen_06ba94794a14f0f0022af6f5f3c16e1e16959d4c.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_4b1eaca3c37a82d19f8dc91f06764170069ca3af.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_91c916e14198f6d18dc89915e379b01070434e91.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_deterministic.hip -> fmha_ck_autogen_8e816fcad5e9ecfca94a6491eb2274bcc41e558b.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16.hip -> fmha_ck_autogen_5fc66c5b53f83bf1e023e81e9d51f0285b3ae731.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2d9c659ba43bb907fd4e3e36a50958288bafd1a3.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask.hip -> fmha_ck_autogen_07ff04fcc273e469737512893ea3fb5876ac131d.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_deterministic.hip -> fmha_ck_autogen_22632f996eb63fbe4bc5748c5897b775087446a0.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16.hip -> fmha_ck_autogen_f5f1797f6b672a55476348571ce17645c8a62869.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_iglp_pssk_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_eee408cf9456ff977aa7d12345e9b2f1e60639f1.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv.hip -> fmha_ck_autogen_303b7b04496e4db7c1ba2436485dc7c8a4c88448.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi.hip -> fmha_ck_autogen_fcb0b08e29b2e1bf181fceceb9dc416e54f52b00.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_deterministic.hip -> fmha_ck_autogen_d06ba4c996570ddab77b6ff1e2a0101b638543eb.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16.hip -> fmha_ck_autogen_fc5ebf0f2200f37ccc0849e0c3745f6e2f00111d.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_dropout_wg16_deterministic.hip -> fmha_ck_autogen_2caba3ab83239e474412fcf89fe0fbef97e51bf1.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_dc184767d723f4995791848cdc68bd948408204f.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_deterministic.hip -> fmha_ck_autogen_c53e295b68e807774ed31bb914e4bc59312a77d7.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16.hip -> fmha_ck_autogen_db0d0cf55d90b3f3c9eecada1db93c420f34b1ae.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_alibi_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_d1c25cfc437d8bd803860e39a45b2f3b9fa48393.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_deterministic.hip -> fmha_ck_autogen_01ca79005067e20e4eed5a72ff9187cde702cd1c.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16.hip -> fmha_ck_autogen_a5e5cae764142683b70d3344cf07dd1edb7d69e2.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_dropout_wg16_deterministic.hip -> fmha_ck_autogen_ca920c3239bb5796b1ab2fc75177eb3b820aa784.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask.hip -> fmha_ck_autogen_806f9ab9baf631df1d3a8d801e4cf93a102526cf.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_deterministic.hip -> fmha_ck_autogen_4b30f472f00bec9da0564ddc40e07112b5f9a117.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16.hip -> fmha_ck_autogen_dc039d422a57c159ea4dbcc867d766ff1b356a07.hip
-fmha_bwd_d64_fp16_group_b32x128x64x32x64x32x32x64x64_r1x4x1_r4x1x1_r1x4x1_w16x16x32_w16x16x16_o1_kr_ktr_vr_psskddv_mask_dropout_wg16_deterministic.hip -> fmha_ck_autogen_5b55946ff3c15a44b9c741e9f6bbbcb5bd4c8577.hip
-fmha_bwd_dot_do_o_d128_bf16_batch_o2.hip -> fmha_ck_autogen_658552954505a2092662071401e135e84956c4c0.hip
-fmha_bwd_dot_do_o_d128_bf16_batch_o2_pdv.hip -> fmha_ck_autogen_53bd60bd2afee49b30a583c32a45ae9f2076db08.hip
-fmha_bwd_dot_do_o_d128_bf16_batch_o2_ps.hip -> fmha_ck_autogen_8e675919a6c7758cbbeecb83b7ac6c62f95cdb46.hip
-fmha_bwd_dot_do_o_d128_bf16_batch_o2_psdv.hip -> fmha_ck_autogen_2d06f77a4054ca615d96636c0e2eba2a89850142.hip
-fmha_bwd_dot_do_o_d128_bf16_group_o2_ps.hip -> fmha_ck_autogen_187963e1969301abfa61d06afc97faea2bb4efb1.hip
-fmha_bwd_dot_do_o_d128_bf16_group_o2_psdv.hip -> fmha_ck_autogen_e7153f9a9b0b7c54ddf2debbe297efcffbb4fcfa.hip
-fmha_bwd_dot_do_o_d128_fp16_batch_o2.hip -> fmha_ck_autogen_3c3b7e4b8c1efe59f79a15512716fce2282a79a7.hip
-fmha_bwd_dot_do_o_d128_fp16_batch_o2_pdv.hip -> fmha_ck_autogen_19cd9f7b08cec83736605af63d9fcaf463a1aea4.hip
-fmha_bwd_dot_do_o_d128_fp16_batch_o2_ps.hip -> fmha_ck_autogen_b4588379eaa268d79fe8f8e4457b009f204a5fb7.hip
-fmha_bwd_dot_do_o_d128_fp16_batch_o2_psdv.hip -> fmha_ck_autogen_23c9b46da8774462de8c24e14b12df3ed596eb57.hip
-fmha_bwd_dot_do_o_d128_fp16_group_o2_ps.hip -> fmha_ck_autogen_5b413bdc825ae863d53dab548f2145dc0de8fd37.hip
-fmha_bwd_dot_do_o_d128_fp16_group_o2_psdv.hip -> fmha_ck_autogen_58a7ab44bbd9fbc97c7805860d5f6ac81d6ae468.hip
-fmha_bwd_dot_do_o_d256_bf16_batch_o2.hip -> fmha_ck_autogen_50f887556a3540609649744957651ca667b91774.hip
-fmha_bwd_dot_do_o_d256_bf16_batch_o2_pdv.hip -> fmha_ck_autogen_eac5952f46f4f2bf06257b00661774eeed48a323.hip
-fmha_bwd_dot_do_o_d256_bf16_batch_o2_ps.hip -> fmha_ck_autogen_efaa0cb33c71cb8ca7b83dd0e7a6c7b01f6b50a9.hip
-fmha_bwd_dot_do_o_d256_bf16_batch_o2_psdv.hip -> fmha_ck_autogen_71e5fb3544dafa9da03fd2de4bb9bd0718f6009f.hip
-fmha_bwd_dot_do_o_d256_bf16_group_o2_ps.hip -> fmha_ck_autogen_3fad30ff0739ab5dede67a96e859f8c474c245f8.hip
-fmha_bwd_dot_do_o_d256_bf16_group_o2_psdv.hip -> fmha_ck_autogen_4bef4d120e71bfcfe61d67aa44d24ceb907c2b9e.hip
-fmha_bwd_dot_do_o_d256_fp16_batch_o2.hip -> fmha_ck_autogen_7d0f767c17385eb7d756cbe8ed444d7cef72dea5.hip
-fmha_bwd_dot_do_o_d256_fp16_batch_o2_pdv.hip -> fmha_ck_autogen_4b68e4d00295b294320b94bc777d7d34609127e0.hip
-fmha_bwd_dot_do_o_d256_fp16_batch_o2_ps.hip -> fmha_ck_autogen_33746071156e9ad46f403a539dc237e0a44122a7.hip
-fmha_bwd_dot_do_o_d256_fp16_batch_o2_psdv.hip -> fmha_ck_autogen_3d45624dc6e33c477c73a155500b015b6c010de8.hip
-fmha_bwd_dot_do_o_d256_fp16_group_o2_ps.hip -> fmha_ck_autogen_8250f27341241086515d833aa53ae873d4ece3fa.hip
-fmha_bwd_dot_do_o_d256_fp16_group_o2_psdv.hip -> fmha_ck_autogen_8793dc3217e154b65ebba065aa10ab4dc2374ae8.hip
-fmha_bwd_dot_do_o_d32_bf16_batch_o2.hip -> fmha_ck_autogen_1a11dd5ebb989503a1c182684e7f247e2f8cd9c2.hip
-fmha_bwd_dot_do_o_d32_bf16_batch_o2_pdv.hip -> fmha_ck_autogen_e16075c3a5fcfe63ba12e854bb1fed6873f014ab.hip
-fmha_bwd_dot_do_o_d32_bf16_batch_o2_ps.hip -> fmha_ck_autogen_937801fbb43fb6797f0425f08d13926b74d87c4a.hip
-fmha_bwd_dot_do_o_d32_bf16_batch_o2_psdv.hip -> fmha_ck_autogen_fecffa403b3631b1957e1a9a06f18fdb3b4eee5f.hip
-fmha_bwd_dot_do_o_d32_bf16_group_o2_ps.hip -> fmha_ck_autogen_5ba578c0e7abf1127dd0370f06d7278656c93ab9.hip
-fmha_bwd_dot_do_o_d32_bf16_group_o2_psdv.hip -> fmha_ck_autogen_345a939a2491166dc520e9a2b9de7e43671e0c2b.hip
-fmha_bwd_dot_do_o_d32_fp16_batch_o2.hip -> fmha_ck_autogen_7393267865f1c2b0aa1a09a586f54cec98eea4ae.hip
-fmha_bwd_dot_do_o_d32_fp16_batch_o2_pdv.hip -> fmha_ck_autogen_93b885d6869400b0dc2ef1b2c2636ddfd21cde31.hip
-fmha_bwd_dot_do_o_d32_fp16_batch_o2_ps.hip -> fmha_ck_autogen_38f8a89468cf9c8606cf12a930db062a83cd0ea0.hip
-fmha_bwd_dot_do_o_d32_fp16_batch_o2_psdv.hip -> fmha_ck_autogen_f974b12e83e214c30995a25631d37df1478927af.hip
-fmha_bwd_dot_do_o_d32_fp16_group_o2_ps.hip -> fmha_ck_autogen_2bb6da1095bd8669c0e48b5cd808cf0dcefa2674.hip
-fmha_bwd_dot_do_o_d32_fp16_group_o2_psdv.hip -> fmha_ck_autogen_0e0a2370f2a320484d8f9f21e3197425c2dbe9ad.hip
-fmha_bwd_dot_do_o_d64_bf16_batch_o2.hip -> fmha_ck_autogen_a9f00f270680de81df7737e848e0408cb070e68b.hip
-fmha_bwd_dot_do_o_d64_bf16_batch_o2_pdv.hip -> fmha_ck_autogen_61220f6dca850a5b5ccf1f619a267c40c37efeca.hip
-fmha_bwd_dot_do_o_d64_bf16_batch_o2_ps.hip -> fmha_ck_autogen_b192c55f002d8540d5f965cc4df0c2e33f4b9ff9.hip
-fmha_bwd_dot_do_o_d64_bf16_batch_o2_psdv.hip -> fmha_ck_autogen_295a523f815eb822d66162d4feb75fe0bc50b648.hip
-fmha_bwd_dot_do_o_d64_bf16_group_o2_ps.hip -> fmha_ck_autogen_292b4f995d622826af5d1f2bffa7ba68467c841a.hip
-fmha_bwd_dot_do_o_d64_bf16_group_o2_psdv.hip -> fmha_ck_autogen_5e840be0741afa4d41fd4789c8300223fdc63ddc.hip
-fmha_bwd_dot_do_o_d64_fp16_batch_o2.hip -> fmha_ck_autogen_0e1dbc9c433ce8ec33ace9e62550261d613db582.hip
-fmha_bwd_dot_do_o_d64_fp16_batch_o2_pdv.hip -> fmha_ck_autogen_6eebd0c2fbfc85f938b10535855c388971129a28.hip
-fmha_bwd_dot_do_o_d64_fp16_batch_o2_ps.hip -> fmha_ck_autogen_0bc7910aac798f0555e9e505ad7f177c9fbbd92c.hip
-fmha_bwd_dot_do_o_d64_fp16_batch_o2_psdv.hip -> fmha_ck_autogen_18b92b4e249195ac3e0c74d246585a4c9e0992fd.hip
-fmha_bwd_dot_do_o_d64_fp16_group_o2_ps.hip -> fmha_ck_autogen_278639d44a4a8372a627a7c31e9527c8faa26f97.hip
-fmha_bwd_dot_do_o_d64_fp16_group_o2_psdv.hip -> fmha_ck_autogen_8e938d0e3ad30db201880642e57758285b2ec4cb.hip
-fmha_fwd_api.hip -> fmha_ck_autogen_1ca3f45d0be2d1119cccd0af042a3e8adeda2ed7.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv.hip -> fmha_ck_autogen_f727911254904ce4341e4ff5f8bafc430b8cfbbf.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi.hip -> fmha_ck_autogen_54208a6e8c5263e38f9ffcb062564ab61d2785ff.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_dropout.hip -> fmha_ck_autogen_1d3ef3d5ded0dfe2a0bafb52ea8f841658db35fd.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse.hip -> fmha_ck_autogen_f15c41ddb04ec7f80235bb3db19198dd6b699713.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse_dropout.hip -> fmha_ck_autogen_a5c4dc0d70c547dbbfb661e879ba7f9adfafc2ea.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_d7290cc4c3036c9205e689cbcc60e7d16b97a7d6.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_dropout.hip -> fmha_ck_autogen_0b2647b5982405a48e8c8888552a4b89386ccdd9.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse.hip -> fmha_ck_autogen_eb278488b2cca114adca5e4614d86f92447f937a.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_29fe68ba10b3480dddc9866c51ca8b5efe962cc3.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_dropout.hip -> fmha_ck_autogen_92992be6252f2afdc368bd4baec4b8a55ae0abf8.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse.hip -> fmha_ck_autogen_501dcf3213efd214cc2ce8c9ba0027f991d241b4.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse_dropout.hip -> fmha_ck_autogen_aa6d13b09f85ee62bb5018608812181fb43afc86.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask.hip -> fmha_ck_autogen_d0f63cafbeb445408c884727b473667fb479675e.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_dropout.hip -> fmha_ck_autogen_7596c14b8fee751d03f42ca48ea4f66e87fc2e2f.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse.hip -> fmha_ck_autogen_c2b719893a4d8a1e71857966d399f06c0a41749c.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse_dropout.hip -> fmha_ck_autogen_071751b1012b90f7b57f8591cd06ae1fd27d9cd3.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_d00f65bc99ca08eba66564d34f72f2769bff9491.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_2273457ac3be01cc1595a015a5f598f8290c77e4.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_63c411351ec59bdbed2590c599f9eddf7807b371.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_042a156e9eb935555ab14a84461959b466c2fb5b.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_eab6cdc59bf216f7045f0cf5f221bb91ec415cd2.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_d703eea8075cacec4d41fee7dc4734f593ee79e8.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_2f32f2d658f1f69840fbad511ce8a3851c859d52.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_e6973d75297bd2c3432a7c88e8a9ee1c9ae693bf.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_854c8003a508ed3f8cbe6967c4ae2635a491c721.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_ceb9544e2a0caae2c9e3dd8bbd2c509e8dca1379.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_e83c604d1b8260958becd1c7c209745ff9151715.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_3b26eafe76cca8e74e819220b6de1f4279d48e43.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_d5e82799f4452e148c3e02acd6526cf30757eb52.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_5435b4651a90e331fcdcf224282457e3dc038a30.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_1573e3d855d28c54af612ab950b081302891d56d.hip
-fmha_fwd_d128_bf16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_4e47f8fa40332c6ed12d9971e0b539049a871c34.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_b285e2f1970b78e18002464eeda63798229bbc3a.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_75f21e38ad01fade35b1db40adabd75eb602410c.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_81f6c575c3fa2ccc7e65022f1ba65c8cfc16541e.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_45b9871c220c0065d74bffeed4021d0304a9625c.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_f028af9e5e3c25800dde938e991aaab4fc1d64aa.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_7fa76fc1b066a15b08dc6c24a7cf33a58b4cb6cb.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_157768cd725813f8111d265cfdfea7f42034e5e9.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_541874a7633e5713720b9d084b6d1c6715a51a17.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_6f88527a2cdb5adf51407f4661a254bb32d7de23.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_a55b47aafc4340e69e300ac61a7601a5c14513b7.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_20d5c3c86398f6ce55abc90db3e362dbf9f457f2.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_8cf1007430da272174d3476d042f398627e83512.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_e7d37e7ee96c392fa24c02a9143438a3a7d05741.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_dc91797c1474a368e9cb056b50b4629d7736c3cb.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_3cce3baac1e3ca03af0c3f4ee4d0158ad1031e9f.hip
-fmha_fwd_d128_bf16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_9d6759d8855c4c6289f1f241a1628cf0406c1b64.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv.hip -> fmha_ck_autogen_b38a1d3cffae01332a3a9d9472ff1b2c443e82af.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi.hip -> fmha_ck_autogen_2cf351fc2c2da4a8e1760a3affc9a5947c6b3bda.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_dropout.hip -> fmha_ck_autogen_bafbef3f13d429ec3e9f4672218998d5669d79f2.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse.hip -> fmha_ck_autogen_3f34433b784d1e405ade3378918641372a30bf6b.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse_dropout.hip -> fmha_ck_autogen_5fb062527121e627871b3f1b2a94b96c42e51205.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_5732094f5917e9164ee0f973ac6ec47245a69101.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_dropout.hip -> fmha_ck_autogen_688aaa193f332ed13e017e78ec07a7c80e45f6c5.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse.hip -> fmha_ck_autogen_1cbf88db44aa5f884438288a325270d29c7a04b6.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_2660282ad39ef034fecbdb74acedfb48620b7dfd.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_dropout.hip -> fmha_ck_autogen_a59423c095db052603d77073d409534bceef425f.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse.hip -> fmha_ck_autogen_3fcc6893456a559c7d22714116022fc69b372266.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse_dropout.hip -> fmha_ck_autogen_c7568e11e44ce70924d27e683190422cfae5c31d.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask.hip -> fmha_ck_autogen_f79def2b4edf6d18f6ef1d6b141f9e0435441f6a.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_dropout.hip -> fmha_ck_autogen_32652a27e8605cef59c8341813b68e7513be23c5.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse.hip -> fmha_ck_autogen_b20e314642cf565e4f32bceffdb5c0e653ab627b.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse_dropout.hip -> fmha_ck_autogen_a74b0e7dd816ad08eec5a1bba6e227afee9813ec.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_a968df29f5ae1463706b7981b3bde55918e1aa65.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_b5248f443a12d96815c04409a00102923c717023.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_291a8bdf9d63b112e7fe5fa7e8835a6789cb8ecf.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_6d5aad18f59e47a3fa3278c7ef1a6372830c33d5.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_c063318cb851ccaa923be12d34c84d839bc64bb8.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_a5a7833f4597bb03a3e845d5580d677e97421040.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_2d9a04b7f41dd6f0db017157a44790f35c626e2d.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_98f5efcd500ce6b9ffc14bc9877e0ba457539925.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_135ea67de101135ed5fe04f5cab1ec1d7b3714bb.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_951343832a5bfd060c8d12da0d8a090f070a717d.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_f24d42e820adc1a26a428d59df7ffdd7f8580176.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_4dbdd9c3f496a27bde68cf86374999ff2dd53505.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_3be7cea6df8e6dd56194e1172f28943667f1c4ef.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_483eaea4096c8f5bee16a64860432f0634a253d8.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_103186dbad604763008e0204a1ea90baecef8877.hip
-fmha_fwd_d128_fp16_batch_shb_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_367e58867c46d96c9bbaa96eaaa9f93595c9e099.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_311104394c8bef8d4ecff35c1409221e723a5a8a.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_976cf509d9c2bf86ba6ee5ded544fa8e6717f590.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_24410fd9a4150c33186a2a365d06d8f6ea621c20.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_b493c99888d82cd2852bfb101f99a2e6a27665b8.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_1fda1c96568eab89a8f6498f8bb23c1223cdc7b0.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_053981d9e7af2ebc0f91e61ac5e25cbe68c95bd8.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_3110540b50e95e99a5cccebe47d9d3a83093c2fb.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_1fcdcb750f382fc7828a9886585f50efbe5be735.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_c3d0eaf9399c863d672e8c08d123739bab837d4b.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_9d69d441f48f9ea346dd8e00376a9a708da3ad87.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_3992d5df4ba2e999caf6889a852db4e1ba078e65.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_f30316cfe49323638f71ba688dd8ff9b2266b335.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_797750ac0b18b48f56ceb4640256e9bd3a36621a.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_942439e4f5644a3a4630481bc7d98834b29b6e1c.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_fac99c3c82b77946f6844699d2333cd532a78a26.hip
-fmha_fwd_d128_fp16_group_hbs_b128x128x32x128x32x128_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_98f9a4f4d85f292b78123599a2e1798f12aa545b.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr.hip -> fmha_ck_autogen_ea591185b1c5f521023e250a26f742984255b241.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi.hip -> fmha_ck_autogen_48300e0aeabe337785d4c7b41796ce65df6cc42a.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_dropout.hip -> fmha_ck_autogen_e514c6b4bc75d95a150104a17972abae77cb47ed.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_lse.hip -> fmha_ck_autogen_a64b4cf3f6706e4b4e0af4402e2263b9a1585f9b.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_lse_dropout.hip -> fmha_ck_autogen_e389d0e4442cd8304081892ddc75043e68a6398c.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask.hip -> fmha_ck_autogen_ab43f4a56c166dad0113f51b337a083f4df7cdb6.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask_dropout.hip -> fmha_ck_autogen_d4645b713821371161a9925dec8a3d6c157ba1aa.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask_lse.hip -> fmha_ck_autogen_0b90a0186d8b8004e3f19886c7992c8e04d0e066.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_d34d6cdcd81a456125ab5e0875466c6334d8e5c8.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_dropout.hip -> fmha_ck_autogen_d0b09e8513646fbb2a007544a63ec9e2b04dc4c2.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_lse.hip -> fmha_ck_autogen_ca3d98ff43fbb80ceb82fc22ab039bee898969b0.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_lse_dropout.hip -> fmha_ck_autogen_7ea9c37d92e344f3cc58cd4d1d00f19167e3623e.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask.hip -> fmha_ck_autogen_db85839ee8d464c5a81b8dad9839f5e0f4b467a8.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask_dropout.hip -> fmha_ck_autogen_32527660fa7aeb9a951a9f2fc3c53989bd141c48.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask_lse.hip -> fmha_ck_autogen_528db08068589c6e4c096054d26a2e5be63285b6.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask_lse_dropout.hip -> fmha_ck_autogen_d600779c17b7b21c18e1308e6d765fe02a7945d3.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv.hip -> fmha_ck_autogen_445e28a8a51cd435130ded2abc9fc606e522c713.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi.hip -> fmha_ck_autogen_8a980749c6b2a18c80426dd189e5506334343ca4.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_da822ea727fb3543e445e4000f7e6ebb946d6a3b.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_f525b59df454ccf53da6cb201e0aa8d09f52a2ad.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_0a2b116fd5065109aae46ee547e4f49ad0e9d6e1.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_366662dccf2f650bcd8123c49006c759cd4c0ef6.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_816c48e129a0235cb3a19124ddb28cce286fb368.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_356f83cb96d0313abcdb24955edd4264df72aed7.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_0e661b5f30566d1f159f060c264849c7ae4772f1.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_dropout.hip -> fmha_ck_autogen_61a9e92183ba87924e73ff0b5e25bd12d6038e69.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse.hip -> fmha_ck_autogen_e502730dea6987e2c038446c448aa08bdcc23113.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_f851da732f397624717160f89271514bc334b59b.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask.hip -> fmha_ck_autogen_fd345632e0cae0d549ba79626a08b1885711deb6.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_937c48d0b7096ad6c8bc445f13f2c8c1934695ab.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_a2482a64659c838f3da55f56e3cbbee1dbfe6722.hip
-fmha_fwd_d256_bf16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_f34fdb8294257d951dcc9c4fa7ecf1192568b91b.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv.hip -> fmha_ck_autogen_0aafb881e34a3794970a1282af740b3f19c138b1.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi.hip -> fmha_ck_autogen_c250ea59ab6e1ee39cce15cbd3f181047cdee31a.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_4ce671f5defd76ca08614a7a1f184c36c0f1e2ab.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_b9627f9c8d0088df0364a64643f2b5dcd951f2bb.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_a6461d72fb6ba50e81de3f661528c96dcfdc3f3c.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_aa82d20635e592edbf00439294835f6f39ad54a3.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_146eb8c40e3146e06936f3141b2c4d92a578ddec.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_c28de8f96c8315877031a2d56261e95fee6aef44.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_39422621a00ff79b2f5ec0dafb957c77693537b3.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_dropout.hip -> fmha_ck_autogen_a0a556c9358ddd6db719458c81d2d6d822a895da.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse.hip -> fmha_ck_autogen_c2fcced07cc194a8050bc7b2f791453b3f5b2064.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_210ef512b7862837f54acbc3b21e135a192647a3.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask.hip -> fmha_ck_autogen_bef3bd014a918feddadc98eed92a7734f9bcd890.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_ae1ab1f4bbe86bb9bbc22e4774648076c321136f.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_52a8a323414448c50571a334f29bc0a38919b61d.hip
-fmha_fwd_d256_bf16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_204a573ce6b7d2f90aede543939315561cc43177.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr.hip -> fmha_ck_autogen_d8901a63986cc28ef24cab012b32114851a8c1ec.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi.hip -> fmha_ck_autogen_12d5c8a4988efe60ef7943ecd73e18a28a736583.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_dropout.hip -> fmha_ck_autogen_e5b65fc519ea7cfcd19f7eddbc3acad6842ff558.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_lse.hip -> fmha_ck_autogen_743176ecb1f0bc800c870861585edf56f88d7739.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_lse_dropout.hip -> fmha_ck_autogen_6b0ef67ce0f178aa2863c4909f5bdd7f766c9b2f.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask.hip -> fmha_ck_autogen_ef40f0acf1885096efb840ec5600ec421c4db331.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask_dropout.hip -> fmha_ck_autogen_523e5bf45ec5008aa3aba4773e68a78e122b2fe7.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask_lse.hip -> fmha_ck_autogen_55cda610c235987e13232e828f8d86fa88030560.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_566b4782793c6526bfce7362efbf6bf069928b2b.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_dropout.hip -> fmha_ck_autogen_cfec97bdfb6fa95e057eaf5a8138853e1c0884f2.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_lse.hip -> fmha_ck_autogen_6905ba47078abd7a5b6a51eb93b26095517e7f70.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_lse_dropout.hip -> fmha_ck_autogen_8840e8899b4e632714632450bcef001c6070f955.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask.hip -> fmha_ck_autogen_d867098db97b3f26e71a151c63b74260bfab21f8.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask_dropout.hip -> fmha_ck_autogen_bc238fd2095b26a167b41cdec8280182330b7b25.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask_lse.hip -> fmha_ck_autogen_b737410b404a51043fc3bd503c0b107c297e4c9f.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_mask_lse_dropout.hip -> fmha_ck_autogen_b4a5715b550f67b8870ba66e1e6282a26cc1dbf3.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv.hip -> fmha_ck_autogen_12207f4b6e7fac27d6c16493a5373f448a2aaae8.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi.hip -> fmha_ck_autogen_7d5667b27f15a06d4040354fba3601d48bb9c045.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_2695783ae8f0034692efd6563f789ef03fd0f4f3.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_60801d21c14796c08377349ec86a6c800af497b7.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_159ee1f1b44d1a8fbaead65d8449413bb616d15e.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_9f0517550c7a23882b95de451e8099ea2186b4ce.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_80f51f0e178c33e6196df1d2e47bd38bf5391cc8.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_489e7be0f85656d012a6451b65f6c1d2613b187d.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_e7de729aa50c10d8101ef504138c3769e3286753.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_dropout.hip -> fmha_ck_autogen_25b3225da1e1842f83592971a1f62a0fe30aa9d3.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse.hip -> fmha_ck_autogen_ce4714e4f33340859c106a3129993e22652262e2.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_bc4e0f0496a34d2fb43c80ce0162ad4183f29064.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask.hip -> fmha_ck_autogen_a9d2be18e2d53a5144f97dfdebb225fcb6d611d3.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_4ab5d6e8fbfd92e9f7e47bda5cfbb0d4162a6319.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_ac1ccde31b47e0e56ee0daab6403fed7895208c7.hip
-fmha_fwd_d256_fp16_batch_shb_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_5cd03e29403ad53d6d52e5e81182ea6ff5aff2be.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv.hip -> fmha_ck_autogen_2005aca3520b171bb82d10ad70fef44f28c19776.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi.hip -> fmha_ck_autogen_c402e84359b2037a29efd1d6ce7213ba7605ab25.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_95061acc6650fc7b79fa1fe5b2b1e083555eec2c.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_1fd9fa7c2e13d0bad5fddb2b5a316bbc09d397ea.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_dd9494d9ac35eba6794a4f9120d2db9932596ef8.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_82d7f61e6313930f063758b61102e7a43b118beb.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_2b50073f6dfeb7ea77d5dce288a1d2f08f8f6362.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_5fa7fafd4227918e0c7f0c6ca3b2bd673cd07279.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_522a2a9435103ed405dc1500d31652f1d431a49d.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_dropout.hip -> fmha_ck_autogen_4b7393d55600c9892558248f4131fc06a6cf3309.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse.hip -> fmha_ck_autogen_d66c30148a6fa816937f2f095802264d3dfa0273.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_8f7166d4bb0c1c9b9999ba16a1adbf09ebfdb6f1.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask.hip -> fmha_ck_autogen_80cf0997573f4bcfbaaf75e40f519580a7495a17.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_48d7d145f96aa8958a9208d0c8887742a8c834fd.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_bb111b7acc269f8d5e70915d3efde4c425aa5f5c.hip
-fmha_fwd_d256_fp16_group_hbs_b128x128x32x256x32x256_r4x1x1_r4x1x1_w32x32x16_qr_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_48435e5dd23e49e19dd313f9891ffec800ce74c2.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv.hip -> fmha_ck_autogen_e2762543d3380185e304f84749a70db1b8d3dd8c.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi.hip -> fmha_ck_autogen_5093976cb7b32a8bd28ce92fc13af00a3e21f737.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_dropout.hip -> fmha_ck_autogen_efc6a7b25710f0626c3af534111b161e1459d2e1.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse.hip -> fmha_ck_autogen_a8a744edfa3a19d1493611df5bd0d4d59b707d43.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse_dropout.hip -> fmha_ck_autogen_e95e3908479965856843317c8b0c42a6961dfd23.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_2b5317b6cde327a842170ebff20c2b03d81379ff.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_dropout.hip -> fmha_ck_autogen_99ae680eed89ea93a3a94586bd5a68dbc5439f37.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse.hip -> fmha_ck_autogen_1edaf9d4270d2ac61c299320e06ba73f44730364.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_0a4e76d89b175e1d9fd2e9fb908d5fce1ebb945d.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_dropout.hip -> fmha_ck_autogen_fba47fa8d9b5375bc408af68b67345ab9dba2eb8.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_lse.hip -> fmha_ck_autogen_830e3532f27b391585d5de90f3bdf97992b67651.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_lse_dropout.hip -> fmha_ck_autogen_66a020f728df204ff51e37d2ddc21afb0aad5e7b.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask.hip -> fmha_ck_autogen_07c3fc96d2bebe546dce6ebf46e5c7a519959599.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask_dropout.hip -> fmha_ck_autogen_74d5f2aef029f2103bb419cc982cae99fd1a9253.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask_lse.hip -> fmha_ck_autogen_58a784fb478ff5b3f1e2da9765a3a777efda92e3.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask_lse_dropout.hip -> fmha_ck_autogen_0766e7aa4b263a811408b285213e47176ee2bdaf.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_bbe23201fbebed25781f249e5c77c31e0e7f9ddb.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_7a890b126da2d8cfbf84f048b779cac2dd56b509.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_58679919fcd292a2a69543de0db94e2985c9d364.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_84fc5e94f89d6a9287cf64662a372784511468dd.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_3bed3aaf24c73073c604a3b23bb4b0358b8e3490.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_dc5ba6d73f331c76e696953606c5b347b6a46f3f.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_b4f12f10d7b968e0d8e7c23f36d3a360de74a905.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_41b68458076e6cb129d3ec793e95b91430a0c8a1.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_56ffe9e21362afe9c3a407c09d5de186954931a6.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_2ba934408c75da5479cc41f96b98ea7d333635ea.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_bd6aa39d0ae3c87d011610cdb5e2e317f337c454.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_aece14f7a220222eb4ce6783ec2b9fce6fde94b8.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_6e240106c771ebea461fc2a87b6da68e510aba70.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_232f61bf31dbb5de5d7039d5ff2338068a759b68.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_e0e48d7edfe9513f24ad9fae68cac3aa940b17dd.hip
-fmha_fwd_d32_bf16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_bc897852a4ca992961843144f4ec4f8b86dd5e9d.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_f1246d1013d954a9316f4432c986d3be9459c548.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_6a4b6226b355bf35d4d07aaef1828091f03ad2ec.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_2b49a9b0801a06dd89c7f7182d7590b515df1592.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_50e7b11019fc2299d70869253877319b03388244.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_7f9bb3486fee7b7c9e24300b8a4e4ce88a11bfc0.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_6785dcec0197fdbb50124ab06efa627f1a2c0567.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_f87991cb7787a29d3ce4711b4ce04c5fb6a14ca9.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_bc4425e30a0b17e8b31726817e8d3177b5c51934.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_54940ce53998becf9bddf56df7d19894a7658168.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_ebb241b947a0adfc8e50c5d71765c14af24593ae.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_3d3f3eb2f5eb1f3287879604892b1c230df85f1d.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_7b9a3bf1a9b37e0bd9bae6249609e5994dc0dba1.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_14221590b90c48d3cf259fb4e834ccfaf7f3209b.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_445cd8fa559588f4264ce6192f2de3e3065365ea.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_7a902ed4ae3cc6558c73b730ff3949778007a230.hip
-fmha_fwd_d32_bf16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_0682150e93f547e00f13cd8984779bf49b91e50c.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv.hip -> fmha_ck_autogen_d86e4dcbe9c4cac8f7c8c5d97ce384ae0cbdbfbc.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi.hip -> fmha_ck_autogen_1df893ee660d37fba7eaca452ae65b3e45a73087.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_dropout.hip -> fmha_ck_autogen_92739f4464512feee083b875e11e11eee4f5b448.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse.hip -> fmha_ck_autogen_65910c8b7a30acc731948ab58467fdbe4fe32f6d.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse_dropout.hip -> fmha_ck_autogen_df5b1c6758d4b8540158299dd0362297083084c2.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_ec7fc24902b1ebd8f2bf8088b0ecf6de8be8362d.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_dropout.hip -> fmha_ck_autogen_9e51083e13aa4dfa8c969f8f916835a8e5e9ca39.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse.hip -> fmha_ck_autogen_b41ea5293bc1c56efa2c4b5681d965aa6f2ce6c3.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_813e60e8405aca3f7fbed19452ae37574ada9a77.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_dropout.hip -> fmha_ck_autogen_0ebacd06455ab20eba78b389462946716b5819f6.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_lse.hip -> fmha_ck_autogen_15b255dde1a9d915e582ee2a83de7d83190c6a24.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_lse_dropout.hip -> fmha_ck_autogen_7b2d3680c3578c7292349b58843aef7a82e0087d.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask.hip -> fmha_ck_autogen_1d21263e16dafe79b9fe2f998847296e575c14e7.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask_dropout.hip -> fmha_ck_autogen_2d23a26e0a59a8323dd97632e610d24624143fbe.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask_lse.hip -> fmha_ck_autogen_4fa883a36a76edb276a66c5d779294f170d6d4b7.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psddv_mask_lse_dropout.hip -> fmha_ck_autogen_9207a63fc55c411c73e4f93306c5ffed800dd249.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_0a68c2f9a3acdd787b81be455cbc7836c8bfd90c.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_4217a48a1677bd26cd48e512f1fc8830a8a551b8.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_2b0bcb241e5a1be1d35366461408d06e095a26ef.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_f3193ea266f3718398bc5622f8bc7042c3527a42.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_bb28a4e95723e3df380f98b5ac107c4df353850b.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_61204f6805d5d830aa6fca2a9b5f238ed63c3a73.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_6649f19deeaea20663bee781af7edced7f7a4fc0.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_d3784fb4c0685d7b651f4113f3c71e050881f3a5.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_ed6bdf67720e938d538a867548ac3579b8238169.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_971a08c2e48d805b295d979b24173a04cf58def0.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_c4997f79435cf64add10506acb97d0647cfbb3d4.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_188a70d526394e254274df95de0727850820326c.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_661b49505cfecbe4ec3e5c7371de3aaaa85ac9d5.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_d63c8c746055851217a514321cd735eaf6937263.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_745705ae121a1a331527cedfe4d31218a428a0df.hip
-fmha_fwd_d32_fp16_batch_shb_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_6fa6478cc27e52fd9511fbff38369c921155cfb9.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_4fa4d21931b9afcbd70b1567995d3eeb6f9308aa.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_d43715cce8935439f90172d141050d78c7e76fb7.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_ae1afeb6cfdf860ff08e4c2f11c922fd5bfa621a.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_f24bd5b92ce6bba640b8ec6b4e53fe35902c5572.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_481415463f0316ebe25ff2fda47c68cc54db3359.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_db5016bff9e5dc37184d2b9417eb351c7ea1c322.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_d64b8b52f4a98801e185e2f132b2f80c29dd0c37.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_09ecb6347009f6a5d5530a6acf90f9f40288cbcf.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_50e59bd079f4d205b613056f975fd2b4e372ab10.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_fd10a3b937e9659716925e39a01d794914b08e26.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_ec51d24ab5f24e003ed6751ae8ae5b327892b15a.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_a5f8b7b2a891aa9f2ab49762eb31d835efdf18b6.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_9a0a70932bd587759df1e5e150b25b0126d7b529.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_9d3d274058bc0a3d4d35d90669587761fdfbdba1.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_23914c00690ac5c4f89cdbbaf00732ba66c5c0ef.hip
-fmha_fwd_d32_fp16_group_hbs_b128x64x16x32x32x32_r2x1x1_r2x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_0befed50a89d80c22b2c8c3d5ba67d73c3d0190e.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv.hip -> fmha_ck_autogen_88c04463f9c5ce565a9daa8c22e16de80fadd707.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi.hip -> fmha_ck_autogen_01e8aedb7b7d77f44a46b2e9b7a826f245aaf4a7.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_dropout.hip -> fmha_ck_autogen_beae876d6da465687f162136231f15767cc7bb14.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse.hip -> fmha_ck_autogen_26f90358e522d7bb7c76c3a2c6010f0f38788bb6.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse_dropout.hip -> fmha_ck_autogen_d7bda8157fb27d544e049fd7d2ec735725f1bf44.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_9fb389d4b5ba590baa951f17da06f0e53d2bfa55.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_dropout.hip -> fmha_ck_autogen_428ce4e14cf94b284ffa735fe03d923cc74c9fe0.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse.hip -> fmha_ck_autogen_900d7f81c73b35ea64095d01c5d48d9190839e0a.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_d2daccc4b3a0f90bff39cb4597f8b7e484613d9e.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_dropout.hip -> fmha_ck_autogen_f280e1639680ac1e5830a21f921bfe2cf364ef42.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse.hip -> fmha_ck_autogen_0dde401aa76cb5425563cbbdb0362748148da3ca.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse_dropout.hip -> fmha_ck_autogen_dc62a8db637d32e7dfdb2521cbdae6e1fbbd5fd1.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask.hip -> fmha_ck_autogen_4cd3de43cc1f7588d62a10362f59d113ee818846.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_dropout.hip -> fmha_ck_autogen_224f9af5e5ca519b21b71a54acb49f50b4999c47.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse.hip -> fmha_ck_autogen_4c8720923c3452e3aebd7b9c1b4b23f0c35d7e4f.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse_dropout.hip -> fmha_ck_autogen_2c7aede7762a524a7a424cc4dc46e43fdedf73a2.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_a98925d99dc484da41dd55700e151cf545cf821d.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_1c65ba6dba01da9caa84ba89453b61d81376763f.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_4b76e5dce9af523422782dd25d8dcf6f25edc68f.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_fe245e9ea974adce2b9807d33b9ba12d916eaffb.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_281d897ad17d7f6db2741b396e6b85a9b8f35286.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_31a968898f0bc6366313e41eddb5e3a3ed12dc98.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_52688999141a72e61322140db29043ef9f7fbc3d.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_92b722cdabcfaa388ccc6ccceb7e42462f3bdcd1.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_47f3ced9b5ddb0dfee8ed5e7df8eca0bbe273047.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_d2dfdb42c1b380e860aa5609302f29698dd27923.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_7fe409f4421193fb48a54aa5f26bd6229d23204c.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_b3a104733f678193068d8642d6560faa03897258.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_df66feebc9a0dcc508ce002c255154622875e524.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_8fa4c40e244b412a07933d369704bcdaa6d5e74c.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_60efa9c427dc278c0d1bc31189f683cd45e4d873.hip
-fmha_fwd_d64_bf16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_8e50ea8dd480012cbe10be392cd26d1870e6ef9b.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_e5ccd5f7ddc894b2717112cbfc766804e02b7bd1.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_4911bdd71351610d55916d452495e599960d0a41.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_d2d08c5470a385d0160b2c1441fd1c30fff1c17c.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_012c0f480917c329f4c3c6c666cf32af2d82b294.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_0bb81407c8a2b3cdc5fecf655b3ad64d5d729cc9.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_7ff65c7abd9b0d8a2df9302d6dc167637b3a72f0.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_d712f23ef88ae5d7b161d36f42d22a5ba53b6354.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_5bc803342862aa30e23e5be7d84e611bc571c529.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_0ace6e29e1d3060c3086c08fe27b471e375f9c75.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_54ff49018f1c12b9fa31e523ad40b9cc162ba34d.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_351425a006aeeff4d69c8570cb6bf1e1427d2c21.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_fcb6ef39c3db49f26f736d6c9221dd825409ec4e.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_f98a6b193fec3203eaa75819f6b51aa45a48f212.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_2d446754d7000673779d15d3e73039fd3c10a720.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_ca00cfdc5592b7440d72482a18781e9cf3afb05a.hip
-fmha_fwd_d64_bf16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_1211733062ed30b876f1d63bffa642d77e258dd6.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv.hip -> fmha_ck_autogen_9b6d08e63b9a90f2524cbfa8c5fcf8b82a1d2d36.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi.hip -> fmha_ck_autogen_e52e3053f30f780f346fa6b7a836ad2554cb85df.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_dropout.hip -> fmha_ck_autogen_3ecf565a5a1c4a09887c67ac3b9a019dca427ac0.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse.hip -> fmha_ck_autogen_52a89981a05963efcea7ba5c1e967638beeebbbb.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_lse_dropout.hip -> fmha_ck_autogen_2173b7c710d418f44dc2b41bec5905024334eae5.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask.hip -> fmha_ck_autogen_b1ad101ce91348266d3885afdf2996a0fdb72135.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_dropout.hip -> fmha_ck_autogen_4da9e9b7277bc90518ab92860bef2097ba96d982.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse.hip -> fmha_ck_autogen_7e1bdde812c332c9fc58613698568a04771b9fa8.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_1acf2f892742b1d236d2b31a8185c6869126adad.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_dropout.hip -> fmha_ck_autogen_155bafb551768855c8c01faa63e44764ebe6c110.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse.hip -> fmha_ck_autogen_f053c9c32518b895daaa3521827f37af78836fb8.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_lse_dropout.hip -> fmha_ck_autogen_adf160741a4f751d2f15d6eb23d4121cdca62b55.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask.hip -> fmha_ck_autogen_34c2db98d8e2e690f499f41cfd5afb831b756f54.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_dropout.hip -> fmha_ck_autogen_0789852b0cd3cc030c78b28f2fd5b6b0546382a4.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse.hip -> fmha_ck_autogen_532a6ffd8a21d3e98342fd401f0247f62ca4e038.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psddv_mask_lse_dropout.hip -> fmha_ck_autogen_d0daa59f5dce6fc3965193ae37d8c82a3d1834e6.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_4a4a00bd6ea27ff20a2903d619e1361b5e27672a.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_93054acb8a9508fd0f0f486367fb62454de47c39.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_b774450ebadaacf23e944aaf8ca90eada01e8a5a.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_2a833fc01e88bd8e256ef64ae8251dd0ed10720b.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_aa522b43c5e5ea69bcabb4c0fe28def2bd081a12.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_67fb736c61088b8dd92fe0371f5c98e23bf9077f.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_b5c3131fb8e5a25bd4a14bc9075eb6fa01b61d02.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_d7fae2c18645d36a181a0bdd2d8ca7a4ac0f6d1d.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_c355189ade9b1a8269230232db754a3881b53168.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_e035773419a9b3631698a3d375d829af55f7731e.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_d992eab7de49033f5480c5e86a69e675db0d2a19.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_5382a30dcf702daae19bd6705864bfe36e09502c.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_167f5328b035ed59a6f05dfee31edd704c4b07ee.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_c1b94e19d762ddc33cc4e94c6675d93cbde21e3d.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_606f5e0b99814b0a82a731de36f28024bc317801.hip
-fmha_fwd_d64_fp16_batch_shb_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_0ad9d68fcee021437e13ffdf94d78252205f5a31.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv.hip -> fmha_ck_autogen_85156f2c556c6ef6180608c361b7b35ede71ffea.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi.hip -> fmha_ck_autogen_890aa875ac13957f00b30210477924697abf0c9e.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_dropout.hip -> fmha_ck_autogen_3108502fd29d3a24b32177bcea968121ee809115.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse.hip -> fmha_ck_autogen_d66b79c4ebdcfd239cecec58203606bc123bd6bb.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_lse_dropout.hip -> fmha_ck_autogen_5efe77ca5c394a60af0313072cdd132216a52bf3.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask.hip -> fmha_ck_autogen_772016803aa3ca6ebe785557118365f9be7c4339.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_dropout.hip -> fmha_ck_autogen_93728d999ae43ee1b5a16e60b90cf8533c7d303f.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse.hip -> fmha_ck_autogen_a1cba1509c413c870c5d784410855ee1bd737da2.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_alibi_mask_lse_dropout.hip -> fmha_ck_autogen_c59ab718fa23f24f09a713ac28a339208a7a5802.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_dropout.hip -> fmha_ck_autogen_afcafd07c1f56e74373ccf37db35976023456d50.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse.hip -> fmha_ck_autogen_ebb9abf5b09e63cbe76390bb46ff7cbefb3141f0.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_lse_dropout.hip -> fmha_ck_autogen_419461cdb5687ebbb7bf0be136071d70420c1619.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask.hip -> fmha_ck_autogen_4beca56234ff6fb4f23b9b24822887fd9a3d0df9.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_dropout.hip -> fmha_ck_autogen_a8a4af070ee46d802cb11086b93daf91538f8a04.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse.hip -> fmha_ck_autogen_79f182ae021e23869d7bebf2a9b4575bdc910ed0.hip
-fmha_fwd_d64_fp16_group_hbs_b128x64x32x64x32x64_r4x1x1_r4x1x1_w32x32x16_qr_async_vr_psskddv_mask_lse_dropout.hip -> fmha_ck_autogen_770ad1eb1b30ad8f1e7c17df486093129b2d5630.hip
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/ck/rename_ck_autogen_files.sh b/aten/src/ATen/native/transformers/hip/flash_attn/ck/rename_ck_autogen_files.sh
deleted file mode 100644
index 0dc441e87ec3e..0000000000000
--- a/aten/src/ATen/native/transformers/hip/flash_attn/ck/rename_ck_autogen_files.sh
+++ /dev/null
@@ -1,11 +0,0 @@
-#!/bin/bash
-set -ex
-
-file_renaming_txt="rename_ck_autogen_files.output.txt"
-rm -rf $file_renaming_txt
-for file in `ls fmha_*wd*hip`; do
-  sha1=$(sha1sum $file | cut -d' ' -f1)
-  new_file="fmha_ck_autogen_${sha1}.hip"
-  mv $file $new_file
-  echo "$file -> $new_file" >> $file_renaming_txt
-done
diff --git a/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.h b/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.h
index ca13a771bb198..f26a914af9052 100644
--- a/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.h
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.h
@@ -143,15 +143,14 @@ mha_fwd_ck(
     const at::Tensor& v, // batch_size x seqlen_k x num_heads_k x head_size
     std::optional<at::Tensor>&
         out_, // batch_size x seqlen_q x num_heads x head_size
-    std::optional<at::Tensor>&
-        alibi_slopes_, // num_heads or batch_size x num_heads
     const float p_dropout,
     const float softmax_scale,
     bool is_causal,
     int window_size_left,
     int window_size_right,
     const bool return_softmax,
-    std::optional<at::Generator> gen_);
+    std::optional<at::Generator> gen_,
+    const std::optional<at::Tensor>& attn_bias_); // batch_size x nheads x seqlen_q x seqlen_k
 
 std::tuple<
     at::Tensor,
@@ -176,7 +175,6 @@ mha_varlen_fwd_ck(
     std::optional<at::Tensor>&
         seqused_k, // b. If given, only this many elements of each batch
                    // element's keys are used.
-    std::optional<at::Tensor>& alibi_slopes_, // num_heads or b x num_heads
     int max_seqlen_q,
     const int max_seqlen_k,
     const float p_dropout,
@@ -186,9 +184,10 @@ mha_varlen_fwd_ck(
     int window_size_left,
     int window_size_right,
     const bool return_softmax,
-    std::optional<at::Generator> gen_);
+    std::optional<at::Generator> gen_,
+    const std::optional<at::Tensor>& attn_bias_);
 
-std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_bwd_ck(
+std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_bwd_ck(
     const at::Tensor& dout, // batch_size x seqlen_q x num_heads, x head_size_og
     const at::Tensor& q, // batch_size x seqlen_q x num_heads x head_size
     const at::Tensor& k, // batch_size x seqlen_k x num_heads_k x head_size
@@ -202,7 +201,9 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_bwd_ck(
     std::optional<at::Tensor>&
         dv_, // batch_size x seqlen_k x num_heads_k x head_size
     std::optional<at::Tensor>&
-        alibi_slopes_, // num_heads or batch_size x num_heads
+        attn_bias_, // batch_size x num_heads x seqlen_q x seqlen_k
+    bool bias_requires_grad,
+    std::optional<at::Tensor>& grad_bias,
     const float p_dropout, // probability to drop
     const float softmax_scale,
     const bool is_causal,
@@ -212,7 +213,7 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_bwd_ck(
     const at::Tensor philox_seed,
     const at::Tensor philox_offset);
 
-std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_varlen_bwd_ck(
+std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_varlen_bwd_ck(
     const at::Tensor& dout, // total_q x num_heads, x head_size
     const at::Tensor&
         q, // total_q x num_heads x head_size, total_q := \sum_{i=0}^{b} s_i
@@ -230,7 +231,9 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_varlen_bwd_ck(
         dv_, // total_k x num_heads_k x head_size, total_k := \sum_{i=0}^{b} s_i
     const at::Tensor& cu_seqlens_q, // b+1
     const at::Tensor& cu_seqlens_k, // b+1
-    std::optional<at::Tensor>& alibi_slopes_, // num_heads or b x num_heads
+    std::optional<at::Tensor>& attn_bias_, // num_heads or b x num_heads
+    bool bias_requires_grad,
+    std::optional<at::Tensor>& grad_bias,
     const int max_seqlen_q,
     const int max_seqlen_k, // max sequence length to choose the kernel
     const float p_dropout, // probability to drop
@@ -272,19 +275,20 @@ mha_fwd(
 #if defined(USE_CK_FLASH_ATTENTION)
   if (at::globalContext().getROCmFAPreferredBackend() ==
       at::ROCmFABackend::Ck) {
+    std::optional<at::Tensor> dummy_attn_bias = std::nullopt;
     return mha_fwd_ck(
         q,
         k,
         v,
         out_,
-        alibi_slopes_,
         p_dropout,
         softmax_scale,
         is_causal,
         window_size_left,
         window_size_right,
         return_softmax,
-        gen_);
+        gen_,
+        dummy_attn_bias); // Not used in flash attention
   } else {
     return mha_fwd_aot(
         q,
@@ -356,6 +360,7 @@ mha_varlen_fwd(
 #if defined(USE_CK_FLASH_ATTENTION)
   if (at::globalContext().getROCmFAPreferredBackend() ==
       at::ROCmFABackend::Ck) {
+    std::optional<at::Tensor> dummy_attn_bias = std::nullopt;
     return mha_varlen_fwd_ck(
         q,
         k,
@@ -364,7 +369,6 @@ mha_varlen_fwd(
         cu_seqlens_q,
         cu_seqlens_k,
         seqused_k,
-        alibi_slopes_,
         max_seqlen_q,
         max_seqlen_k,
         p_dropout,
@@ -374,7 +378,8 @@ mha_varlen_fwd(
         window_size_left,
         window_size_right,
         return_softmax,
-        gen_);
+        gen_,
+        dummy_attn_bias); // Not used in flash attention
   } else {
     return mha_varlen_fwd_aot(
         q,
@@ -447,25 +452,34 @@ inline std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_bwd(
 #if defined(USE_CK_FLASH_ATTENTION)
   if (at::globalContext().getROCmFAPreferredBackend() ==
       at::ROCmFABackend::Ck) {
-    return mha_bwd_ck(
-        dout,
-        q,
-        k,
-        v,
-        out,
-        softmax_lse,
-        dq_,
-        dk_,
-        dv_,
-        alibi_slopes_,
-        p_dropout,
-        softmax_scale,
-        is_causal,
-        window_size_left,
-        window_size_right,
-        deterministic,
-        philox_seed,
-        philox_offset);
+    std::optional<at::Tensor> non_null_dbias = std::nullopt;
+    auto[dQuery,
+         dKey,
+         dValue,
+         dSoftmax,
+         dBias] = mha_bwd_ck(
+                             dout,
+                             q,
+                             k,
+                             v,
+                             out,
+                             softmax_lse,
+                             dq_,
+                             dk_,
+                             dv_,
+                             alibi_slopes_,
+                             false,              // bias_requires_grad
+                             non_null_dbias,
+                             p_dropout,
+                             softmax_scale,
+                             is_causal,
+                             window_size_left,
+                             window_size_right,
+                             deterministic,
+                             philox_seed,
+                             philox_offset);
+    // for FA return [dQ, dV, dK, dSoftmax]
+    return std::make_tuple(std::move(dQuery), std::move(dKey), std::move(dValue), std::move(dSoftmax));
   } else {
     return mha_bwd_aot(
         dout,
@@ -488,6 +502,10 @@ inline std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_bwd(
         philox_offset);
   }
 #else
+  if(at::globalContext().getROCmFAPreferredBackend() ==
+    at::ROCmFABackend::Ck) {
+    TORCH_WARN_ONCE("Warning! You have opted to use CK flash attention backend in a build that was not compiled using USE_CK_FLASH_ATTENTION=1. Please set this variable and try again. Defaulting to use aotriton backend...");
+  }
   return mha_bwd_aot(
       dout,
       q,
@@ -543,30 +561,39 @@ inline std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_varlen_bwd
 #if defined(USE_CK_FLASH_ATTENTION)
   if (at::globalContext().getROCmFAPreferredBackend() ==
       at::ROCmFABackend::Ck) {
-    return mha_varlen_bwd_ck(
-        dout,
-        q,
-        k,
-        v,
-        out,
-        softmax_lse,
-        dq_,
-        dk_,
-        dv_,
-        cu_seqlens_q,
-        cu_seqlens_k,
-        alibi_slopes_,
-        max_seqlen_q,
-        max_seqlen_k,
-        p_dropout,
-        softmax_scale,
-        zero_tensors,
-        is_causal,
-        window_size_left,
-        window_size_right,
-        deterministic,
-        philox_seed,
-        philox_offset);
+    std::optional<at::Tensor> non_null_dbias = std::nullopt;
+    auto[dQuery,
+         dKey,
+         dValue,
+         dSoftmax,
+         dBias] = mha_varlen_bwd_ck(
+                                    dout,
+                                    q,
+                                    k,
+                                    v,
+                                    out,
+                                    softmax_lse,
+                                    dq_,
+                                    dk_,
+                                    dv_,
+                                    cu_seqlens_q,
+                                    cu_seqlens_k,
+                                    alibi_slopes_,
+                                    false,          // bias_requires_grad
+                                    non_null_dbias,
+                                    max_seqlen_q,
+                                    max_seqlen_k,
+                                    p_dropout,
+                                    softmax_scale,
+                                    zero_tensors,
+                                    is_causal,
+                                    window_size_left,
+                                    window_size_right,
+                                    deterministic,
+                                    philox_seed,
+                                    philox_offset);
+    // for FA return [dQ, dV, dK, dSoftmax]
+    return std::make_tuple(std::move(dQuery), std::move(dKey), std::move(dValue), std::move(dSoftmax));
   } else {
     return mha_varlen_bwd_aot(
         dout,
@@ -620,5 +647,36 @@ inline std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_varlen_bwd
       philox_offset);
 #endif
 }
+/*
+std::tuple<
+    at::Tensor, // output
+    at::Tensor, // q
+    at::Tensor, // k
+    at::Tensor, // v
+    at::Tensor, // lse
+    at::Tensor, // seed
+    at::Tensor, // offset
+    at::Tensor> // dropout randval
+mem_eff_forward_ck(
+    const at::Tensor& q,
+    const at::Tensor& k,
+    const at::Tensor& v,
+    float p_dropout,
+    const bool return_dropout_randval,
+    std::optional<bool> is_causal,
+    std::optional<double> scale,
+    const std::optional<at::Tensor>& attn_bias_,
+    std::optional<at::Tensor>& out_,
+    const std::optional<at::Tensor>& cu_seqlens_q,
+    const std::optional<at::Tensor>& cu_seqlens_k,
+    const std::optional<at::Tensor>& seqstart_q,
+    const std::optional<at::Tensor>& seqstart_k,
+    std::optional<at::Generator> gen_,
+    std::optional<at::Tensor>& seqused_k_,
+    std::optional<at::Tensor>& alibi_slopes_
+);
+*/
+
+
 
 } // namespace pytorch_flash
diff --git a/test/test_transformers.py b/test/test_transformers.py
index af711a6fb67ea..3bc7a6fe75868 100644
--- a/test/test_transformers.py
+++ b/test/test_transformers.py
@@ -2599,10 +2599,20 @@ def test_attention(backend: SDPBackend, permute_order: list[list[int]]):
     @unittest.skipIf(not PLATFORM_SUPPORTS_MEM_EFF_ATTENTION, "Fused SDPA was not built for this system")
     @parametrize("mask_dim", [1, 2, 3, 4])
     def test_mem_efficient_attention_mask_variants(self, device, mask_dim: list[int]):
+        torch.backends.cuda.preferred_rocm_fa_library("ck")
         dtype = torch.float16
         make_tensor = partial(torch.rand, device=device, dtype=dtype, requires_grad=True)
         batch, num_heads, head_dim = 8, 8, 64
         seq_len_q, seq_len_kv = 64, 15
+
+        #batch, num_heads, head_dim = 1, 4, 8
+        #seq_len_q, seq_len_kv = 16, 32
+        print("")
+        print("batch     : " , batch)
+        print("nheads    : " , num_heads)
+        print("hdim      : " , head_dim)
+        print("seqlen_q  : " , seq_len_q)
+        print("seqlen_kv : " , seq_len_kv)
         query = make_tensor(SdpaShape(batch, num_heads, seq_len_q, head_dim))
         kv_shape = SdpaShape(batch, num_heads, seq_len_kv, head_dim)
         key, value = make_tensor(kv_shape), make_tensor(kv_shape)
@@ -2620,8 +2630,9 @@ def test_mem_efficient_attention_mask_variants(self, device, mask_dim: list[int]
         out.sum().backward()
 
     @unittest.skipIf(not PLATFORM_SUPPORTS_MEM_EFF_ATTENTION, "Fused SDPA was not built for this system")
-    @parametrize("dtype", [torch.float, torch.float16])
+    @parametrize("dtype", [torch.float16])
     def test_mem_eff_attention_non_contiguous_mask(self, device, dtype):
+        torch.backends.cuda.preferred_rocm_fa_library("ck")
         make_tensor = partial(torch.rand, device=device, dtype=dtype, requires_grad=True)
         batch, num_heads, head_dim = 8, 8, 64
         seq_len_q, seq_len_kv = 64, 16
@@ -2635,8 +2646,9 @@ def test_mem_eff_attention_non_contiguous_mask(self, device, dtype):
         out.sum().backward()
 
     @unittest.skipIf(not PLATFORM_SUPPORTS_MEM_EFF_ATTENTION, "Fused SDPA was not built for this system")
-    @parametrize("dtype", [torch.float, torch.float16])
+    @parametrize("dtype", [torch.float16])
     def test_mem_eff_attention_long_sequence_mask(self, device, dtype):
+        torch.backends.cuda.preferred_rocm_fa_library("ck")
         if torch.cuda.get_device_properties('cuda').total_memory < 80 * 2**30:
             unittest.skip("This test requires substatnial GPU memory.")
             return
@@ -2694,11 +2706,13 @@ def test_singelton_head_dim_stride_ne_1(self, device):
             scaled_dot_product_attention(query, key, value)
 
     @unittest.skipIf(not PLATFORM_SUPPORTS_MEM_EFF_ATTENTION, "Fused SDPA was not built for this system")
-    @parametrize("type", ["dense", "nested"])
+    #@parametrize("type", ["dense", "nested"])
+    @parametrize("type", ["nested"])
     @parametrize("is_contiguous", [True, False])
     def test_scaled_dot_product_attention_fused_kernels_packed(self, device, type: str, is_contiguous: bool):
+        torch.backends.cuda.preferred_rocm_fa_library("ck")
         make_tensor = partial(rand_sdpa_tensor, type=type, device=device, dtype=torch.float16, packed=True)
-
+        
         batch_size, seq_len, num_heads, head_dim = 32, 64, 16, 64
         shape = SdpaShape(batch_size, num_heads, seq_len, head_dim)
 
@@ -2709,7 +2723,9 @@ def test_scaled_dot_product_attention_fused_kernels_packed(self, device, type: s
         query = query.view(batch_size, -1, num_heads, head_dim).transpose(1, 2)
         value = value.view(batch_size, -1, num_heads, head_dim).transpose(1, 2)
         key = key.view(batch_size, -1, num_heads, head_dim).transpose(1, 2)
-
+        #print("python_q shape: ", query.size(0))
+        #print("python_v shape: ", value.size(1))
+        #print("python_k shape: ", key.size(5))
         if is_contiguous:
             query = query.contiguous()
             key = key.contiguous()
@@ -2726,10 +2742,11 @@ def test_scaled_dot_product_attention_fused_kernels_packed(self, device, type: s
         self.assertEqual(actual.contiguous(), math_ref.contiguous(), atol=2e-3, rtol=1e-2)
 
     @unittest.skipIf(not PLATFORM_SUPPORTS_FUSED_ATTENTION, "Fused SDPA was not built for this system")
-    @parametrize("type", ["dense", "nested"])
+    @parametrize("type", ["dense"])
     @parametrize("fused_kernel", [SDPBackend.FLASH_ATTENTION, SDPBackend.EFFICIENT_ATTENTION] if
                  PLATFORM_SUPPORTS_FLASH_ATTENTION else [SDPBackend.EFFICIENT_ATTENTION])
     def test_scaled_dot_product_attention_fused_kernels_packed_accuracy(self, device, type: str, fused_kernel: str):
+        torch.backends.cuda.preferred_rocm_fa_library("ck")
         def rand_nt(shape):
             batch, seq_len, num_heads, head_dim = shape
             tensors = [6 * torch.rand((seq_len, 3 * num_heads * head_dim), device=device, dtype=torch.float32) - 3
@@ -2794,12 +2811,14 @@ def rand_tensor(shape):
     @parametrize("contiguous_inputs", [True, False])
     @parametrize("is_causal", [True, False])
     def test_sdp_mem_efficient_grad_against_math(self, device, contiguous_inputs: bool, is_causal: bool):
+        torch.set_printoptions(profile="full")
+        torch.backends.cuda.preferred_rocm_fa_library("ck")
         batch_size, seq_len, num_heads, head_dim = 4, 4, 2, 16
         make_tensor = partial(rand_sdpa_tensor, type="dense", device=device,
-                              dtype=torch.float64, requires_grad=True, packed=True)
+                              dtype=torch.float16, requires_grad=True, packed=True)
 
         qkv = make_tensor(SdpaShape(batch_size, num_heads, seq_len, head_dim))
-        qkv_lp = qkv.detach().clone().to(torch.float32).requires_grad_()
+        qkv_lp = qkv.detach().clone().to(torch.float16).requires_grad_()
 
         query, key, value = qkv.chunk(3, dim=-1)
         query_lp, key_lp, value_lp = qkv_lp.chunk(3, dim=-1)
@@ -2829,13 +2848,19 @@ def test_sdp_mem_efficient_grad_against_math(self, device, contiguous_inputs: bo
                 query_lp, key_lp, value_lp, None, 0.0, is_causal)
 
         rand_upward = torch.rand_like(out)
-        rand_upward_lp = rand_upward.to(torch.float32)
+        rand_upward_lp = rand_upward.to(torch.float16)
 
         out.backward(rand_upward)
         out_lp.backward(rand_upward_lp)
 
         # Cast up and compare
-        self.assertEqual(qkv.grad, qkv_lp.grad.to(torch.float64), atol=1e-5, rtol=1e-5)
+        #print(out)
+        print(qkv.grad)
+        print("=================================================================")
+        print(qkv_lp.grad)
+        #print(out_lp)
+        self.assertEqual(qkv.grad, qkv_lp.grad.to(torch.float16), atol=1e-5, rtol=1e-5)
+        #self.assertEqual(qkv, qkv_lp.to(torch.float16), atol=1e-5, rtol=1e-5)
 
     @unittest.skipIf(not PLATFORM_SUPPORTS_FLASH_ATTENTION, "Flash Attention was not built for this system")
     @parametrize("contiguous_inputs", [True, False])
diff --git a/third_party/composable_kernel b/third_party/composable_kernel
index 50ee4267e27b8..8086bbe3a78d9 160000
--- a/third_party/composable_kernel
+++ b/third_party/composable_kernel
@@ -1 +1 @@
-Subproject commit 50ee4267e27b875d149e642f4cebd47be1dc3b57
+Subproject commit 8086bbe3a78d931eb96fe12fdc014082e18d18d3